Reflex: estimation, control, gaits, skills, shields
17,999 words38 subsections24 tables7 listingsPlan version 1.1
On this page38
22.1 What it owns
Reflex owns every closed loop that needs no learning, and the safety layers that wrap any loop, learned or not: the state estimator, the baseline controller synthesized from a model, the sampling-based predictive controller, the gait analysis, skill artifacts, the energy-budget safety tier and the runtime shields.
| Crate | Layer | Owns |
|---|---|---|
frob-control |
4 | The invariant filter; controller synthesis by morphology class; control programs over the shared cone solver; the residual wrapper and the ladder; the sampling controller; skill artifacts and their witnesses; backup controllers |
frob-gait |
4 | Periodic orbits of a closed loop, Floquet stability, continuation in design parameters, design gradients (analysis A-04 made quantitative) |
frob-shield |
5 | The energy budget and its monitor; barrier filters with their optimality check; assumption monitors; margin calibration |
DESIGN: the traits Controller, ActionFilter and StateSource are declared beside the step plan in frob-dyn
(layer 3). frob-gait, frob-learn and frob-shield accept any implementation of them, so no crate in layer 4
needs an edge to another crate in layer 4 (rule L-1). The enumeration of gait patterns is finite group theory and
lives in frob-algebra; frob-control reads phase tables from there, and frob-gait owns the orbits.
Why it exists. Training a policy for every candidate design costs minutes to about an hour per candidate. A published quadruped joystick policy trains in about 410 s on one RTX 4090, humanoid policies train in under 30 minutes on two, and published agent-driven design loops spend about an hour of reinforcement learning per round (FACT, Appendix H). Reflex synthesizes an estimator and a baseline controller from the model in seconds, so a new design can be simulated at once (GOAL-13). The baseline is for feedback, not for peak performance.
What that is worth, and what is untested. If a baseline controller can stand in for training, the design cycle of §28 is about 3.0 to 3.5 times shorter than an incumbent quadruped cycle and about 15 times shorter than a humanoid one (FACT: arithmetic on the published figures, Appendix H). Whether it can stand in is a HYPOTHESIS. GATE-REFLEX-02 asks only that designs traverse five metres; it does not ask whether the baseline ranks designs as trained policies do. GATE-REFLEX-04 asks that, per morphology class, with consequences fixed in advance.
What Reflex does not claim. No controller here is called deployable. A controller that works in the model is
a Tested object in the model. Reality is a separate claim (§8, §30).
22.2 Decisions
| ADR | Decision | Reason |
|---|---|---|
| ADR-REFLEX-01 | Every design gets an inspectable, hand-specified baseline before any learned component. A learned residual may be added on top; it never replaces the baseline silently. | A failure can then be assigned to the body, the baseline or the residual. An end-to-end policy cannot say which. |
| ADR-REFLEX-02 | Control programs, force distribution, skill allocation and shield projection all use the one symmetric-cone solver of frob-algebra, with a fixed iteration count and a mandatory optimality check. |
One kernel family serves physics and control (C-06, C-13), and one dual-certificate format serves every infeasibility report. |
| ADR-REFLEX-03 | The sampling-based predictive controller is a screening rung, an out-of-distribution fallback and a demonstration source. It never stands in for a trained policy as the evaluator of record. Where it is the delivered controller, its own episodes are admitted under a frozen planner record (§22.8). | Cost arithmetic in §22.8: a few hundred episodes under it cost more than training once, and its plant equals its model. |
| ADR-REFLEX-04 | Gait orbits are computed for the design under its synthesized baseline controller, first on the compliant tier by differentiating the discrete step. Event-located orbits with saltation matrices are the referee route. | An orbit belongs to a closed loop, not to a body. The compliant step map is continuous, so the first implementation needs no event bookkeeping. |
| ADR-REFLEX-05 | Simultaneous multi-foot impacts are handled, not refused. | Trot, bound and pronk have simultaneous touchdowns by symmetry. Refusing them refuses the gaits the analysis exists to find. |
| ADR-REFLEX-06 | Safety is three separate mechanisms with separate claims: an energy budget that needs no contact model, runtime filters classed Tested, and reduced-order barriers with a measured margin. No full-order certificate is attempted. |
Published sum-of-squares certificates stop at about five to ten states; the template-to-full error cannot be computed in advance (FACT, Appendix H). |
| ADR-REFLEX-07 | A symmetric controller is tied only under the intersection of the design's group with the stabilizers of payload, gravity, boundary conditions, task, observations and current contact mode. | A group that the closed loop does not have constrains the controller to something false (§22.13). |
| ADR-REFLEX-08 | Every command that reaches a driver records which layer produced it, and every fallback is an event in the telemetry. | "It worked" is not a result if the backup controller did the work. |
22.3 The estimator S
A contact-aided invariant extended Kalman filter on SE₂₊ₖ(3) with IMU biases: inputs are the IMU, the encoders
and the contact switches; outputs are base pose and velocity with covariance (C-13). Wheeled bases use wheel
odometry with the IMU on the same filter structure. Every evaluation can run with --state estimated or
--state privileged, and the difference is reported, because a controller that only works on privileged state
is not a result.
DESIGN additions, each a correctness condition for the sentence above:
- Declared error convention. The filter manifest states whether its error is left- or right-invariant and carries the covariance transport between the two, Σ_right = Ad_X̂·Σ_left·Ad_X̂ᵀ to first order. Biases, delays and uncertain contact add terms that are not group-affine, so log-linearity is evaluated, not assumed: the filter reports the measured departure from the linear error equation on the standard battery.
- Gauge freedom. Under proprioception and gravity, absolute position and yaw are unobservable: a subspace of dimension four. The filter's linearization must have exactly that null space at every estimate (PO-REFLEX-01), checked analytically and by finite perturbation of the true initial state along those directions.
- Contact handling. A contact point is added to the state at touchdown and marginalized at liftoff. A contact whose slip estimate exceeds its threshold is down-weighted, never trusted and never silently dropped.
- Namespaces. Truth, measurements, belief and policy observations are separate namespaces with declared access (§20). An out-of-sequence measurement may update the filter's future outputs; decisions already taken and the observation actually delivered at each decision instant are preserved in the record.
- Scoring rule. Estimator error is scored after one translation and one yaw alignment per trajectory, and never after a fitted scale. Counterexample: an estimate of 2 m for a true motion of 1 m has zero residual after a fitted factor of ½. A scale fit is refused (FROB-REF-464).
- A comparator. A conventional quaternion extended Kalman filter runs on the same packets. Normalized
innovations, covariance coverage and aligned drift are reported for both. A filter that reports small
uncertainty while drifting sets the flag
overconfident, which is a failure distinct from drift.
Verb: control.estimate replays a run's packets through the filter and its comparator. Gate: GATE-REFLEX-03.
22.4 Baseline controller synthesis by morphology class
frob control synth <design> detects the morphology class from the model and assembles a controller. The last
two columns are the ladder of §22.7: what to try next for the class, and the evidence that must exist first.
| Class | Baseline controller | Tier | Next control experiment | Evidence required first |
|---|---|---|---|---|
| Legged | A gait from the atlas (C-10) chosen by leg count and target Froude number; footstep placement by a capture-point heuristic; stance forces by a single-step force-distribution program over friction cones; swing by Cartesian PD through the leg IK kernel; joint PD at the driver | S | Whole-body inverse dynamics plus a learned residual for trot | Foot slip, friction margin, recovery after a payload shift |
| Legged, predictive | Centroidal single-rigid-body model-predictive control over the cones of C-06 | F | Contact-wrench feasibility with whole-body tracking | Momentum balance, step recovery, actuator margins |
| Multi-legged (six or more) | A phase-oscillator network on the atlas's phase lags; a support-aware wave or tripod schedule; the same stance program | S | A symmetry-reduced gait family, and adaptation with one leg disabled | Static support margin at rest; fall and energy rates under disturbance |
| Biped balance | LQR on the linearization about a stand, then the legged controller. Standing balance and a supported weight shift come before walking. | F | Centroidal planning with foot-wrench feasibility | Momentum balance, foot-wrench feasibility, step recovery |
| Wheeled | Wheel speeds from the commanded body twist through the kinematic connection (C-07), traction-limited acceleration, path tracking through flat outputs; wheel loops saturate at the card's current limit | S | Slip-aware odometry and receding-horizon docking | Circle tracking in the ideal model, then torque-limited acceleration and braking |
| Wheeled, steered | A curvature command converted to steering angles and paired wheel rates | S | Low-speed nonholonomic planning with steering lag | A turn-radius sweep including reverse and saturation |
| Omnidirectional | Body twist to wheel speeds through the connection, after the growth-vector rank check of A-07 | S | Traction-aware allocation under roller slip | Lateral tracking; a planted roller misalignment found as a rank drop |
| Tracked | Skid-steer with a slip estimate | S | Torque allocation that minimizes scrub energy and heading loss | Pivot work on two surfaces against an independent model |
| Hybrid | Roll where the route graph allows, step where it does not; an explicit drive, stand and step mode graph with readiness guards | F | Transitions selected by predicted energy, slip and clearance | Deployment under load; impulse and completion time across a threshold |
| Arm on a base | Operational-space control with the base as a floating frame | F | Payload-dependent whole-body planning and impedance | Grasp retention, base tip margin, task-space error on a moving target |
Gains come from the model: joint-space natural frequencies from the linearization and the reflected inertia, not from a table.
DESIGN rules that hold for every row:
- Errors live on the group. Pose error is e = log(T_d⁻¹·T), and the desired twist is expressed in the same tangent frame as the measured twist. Near a rotation error of π the controller re-charts by a declared rule; a derivative requested at the cut receives a typed refusal, never a silent sign flip.
- The whole-body program. Unknowns are joint accelerations, contact wrenches and joint torques. Constraints are the equations of motion, the four-dimensional contact cone of §18 (normal, two tangential, spin), the actuator box read from the card's torque-speed envelope at the current speed, and joint limits. The objective is a prioritized list of task residuals with slack.
- Infeasibility is an answer. An infeasible program returns its active inequalities, its residuals, a dual certificate (§22.5) and the prescribed fallback, which is the last feasible command or the backup controller of §22.12. The task records every fallback (INV-REFLEX-01).
- Actuator bounds are not negotiable. The driver model of §15 enforces the envelope in the plant. A run that
succeeded only while the controller requested more than the continuous envelope for longer than the card's
duty rating is marked
actuator-infeasibleand counts as a failure of the engineering constraint. - Schedules are proposals. A relaxed or smoothed contact schedule is re-checked under the committed contact law before it is used.
- Identity. A
ControllerIdis the content hash of the class, the derived gains, the programs and their iteration counts, the planning configuration, the compute-latency model (the delay from decision to command, in the clock contract of §20) and the symmetry group actually used (INV-REFLEX-02).
22.5 One cone solver S
The force-distribution and predictive programs are small dense conic programs (tens to a few hundred variables).
They are solved by the symmetric-cone solver that frob-algebra provides to the contact solve, compiled as
kernels so the baseline controller runs batched on any tier.
DESIGN. The solver runs a fixed number of iterations in lockstep across worlds and ends in one of three states per world, which the caller must inspect:
| State | Meaning | What is returned |
|---|---|---|
optimal |
Primal feasibility, dual feasibility, complementarity and stationarity all within tolerance | The solution and the four residuals |
infeasible |
A dual certificate was found | A vector y with Aᵀy in the dual cone K* and bᵀy < 0 |
unconverged |
Neither of the above within the iteration budget | The last iterate, marked unusable; the caller takes its fallback |
For a program "find x in K with A·x = b", a vector y with Aᵀy ∈ K* and bᵀy < 0 proves that no solution exists,
because 0 ≤ ⟨Aᵀy, x⟩ = yᵀb < 0 would follow. The nonnegative orthant and the second-order cone are self-dual, and
the four-dimensional contact cone is a second-order cone after scaling the tangential and spin components by
the friction coefficient and the effective patch radius. The certificate is rounded to rationals and re-checked
exactly (PO-REFLEX-03), which is what makes an infeasibility report class Certified for the stated program.
Smallest case, by hand. A body of weight 50 N must be pushed along level ground with a horizontal force of 30 N through one contact with μ = 0.5. Unknown x = (fₙ, fₜ) with |fₜ| ≤ μ·fₙ, and the task fixes x = (50, 30). Take y = (0.5, −1). Then y lies in the dual cone {(a, c): a ≥ μ·|c|}, and yᵀb = 25 − 30 = −5 < 0. The certificate reads as physics: every admissible contact force has 0.5·fₙ − fₜ ≥ 0, and the task needs −5 N.
22.6 Reporting: three outcomes, never a bare failure S
control synth returns one of three outcomes, never a bare failure: not-synthesizable with the reason (for
example no gait in the atlas satisfies static stability with these leg counts), synthesized-failed with the
failure taxonomy of §29, or ok. A design that fails under its baseline is not thereby a bad design, and the
report says that.
DESIGN additions:
not-synthesizablecarries a witness whenever the reason is a program: the dual certificate of §22.5 for a force or torque demand that the body cannot meet, with the binding actuator or contact named.synthesized-failedcarries the fallback log, the first layer that failed, and the next rung of the ladder with the evidence that rung requires.okcarries the saturation report, the fraction of ticks spent in each layer, and the standing of the result: a baseline score isscreeninguntil GATE-REFLEX-04 has passed for the design's class and task family.
22.7 An inspectable baseline plus a learned residual F
What it is. The command is u = limit(u_base(x̂, t) + Δ(o)), where u_base is the synthesized controller, Δ is
a small learned network trained by Cortex (§23), and limit is the actuator limiter. The residual has bounded
authority: |Δⱼ| ≤ ρⱼ, a declared fraction of each channel's range (DESIGN default ρ = 0.3, decided by
OPEN-REFLEX-3). Its output layer starts at zero, so training starts from the baseline's behaviour.
What it must satisfy.
- With Δ ≡ 0 the wrapped controller reproduces the baseline's digests bit for bit in class D2 (PO-REFLEX-02).
- Each tick records its layer:
baseline,baseline+residual,shield-filtered,fallback:<reason>orbackup. A residual channel pinned at its bound for more than a declared fraction of ticks sets the flagresidual-saturated: the baseline is then not the controller, and the result says so. - Two comparisons are kept apart when bodies are compared (§28). A shared controller isolates the change in the body. Controllers retuned with equal resources compare the best systems. A report states which one it is.
The ladder. Each rung is tried only when the rung below has failed for a recorded reason that is not a design refusal, and each rung names the evidence it needs.
| Rung | Controller | Cost per design (TARGET) | Escalate when |
|---|---|---|---|
| R0 | Synthesized baseline | Under 5 s of CPU | synthesized-failed with a control cause in the taxonomy |
| R1 | Baseline with a handful of gains tuned by derivative-free search | Under 2 min on a fast tier | Tracking or recovery limits remain after tuning |
| R2 | Baseline plus bounded residual | At most 20% of a specialist's training steps | residual-saturated, or the class has no baseline |
| R3 | Whole-body inverse dynamics plus residual | As R2 | The task needs contact sequences the baseline cannot schedule |
| R4 | Specialist policy trained end to end, or a family policy fine-tuned (§23.11) | One training run | Finalists only (§28) |
Verb: control.ladder returns the rung a design is on, the reason, and the next experiment with its required
evidence. Gate: GATE-REFLEX-05. Kill gate: KILL-RESIDUAL.
22.8 The sampling-based predictive controller F
What it is. Controller kind mpc-sample: at each control tick it samples perturbed command sequences, rolls
each out on the design's own compiled model inside the batch, and applies the first command of a weighted
average. It needs no training and no gait. It is prior art on incumbent simulators, including real-time control
of a real quadruped from about 30 samples on a CPU (FACT, Appendix H), so it is no differentiator. It is here
because two roles need it.
Algorithm (DESIGN).
- The decision variable is a set of spline knots of joint targets over a horizon T_h, passed through the actuator model of §15. Defaults: T_h between 0.4 and 0.9 s, control at 50 Hz.
- For each of N_it annealing iterations, draw N_s perturbations of the knots from counter-based streams keyed by (world address, tick, iteration, sample), so a planner run replays exactly.
K-forkcopies the world's full state, warm starts and queues included, into N_s planning worlds.- Roll each planning world over the horizon on a planning configuration: a coarser step, fewer Newton
iterations, fewer contact samples. The configuration is part of the
ControllerId. K-costscores each rollout with morphology-independent terms: body-twist tracking, attitude error on the group, a centre-of-mass height band, electrical power, friction utilization above 0.8, and limit barriers.K-weightforms softmax weights, updates the knots, shrinks the noise, and the loop repeats.
The three kernels are requirements on the kernel library (§16, §21). Planning worlds are forks: the digest of the controlled world must not depend on how many forks ran beside it (the batch-independence test of §35).
Cost arithmetic. Let h_p be the planning step and H = T_h / h_p. One tick costs N_s·H·N_it model steps.
| Configuration (FACT, Appendix H) | N_s | H | N_it | Steps per tick | Steps per 20 s episode at 50 Hz |
|---|---|---|---|---|---|
| Published quadruped trot, 20 ms planning step | 2,048 | 16 | 2 | 65,536 | 6.6 × 10⁷ |
| Published humanoid jog, 20 ms planning step | 2,048 | 25 | 4 | 204,800 | 2.0 × 10⁸ |
| Lowest published sample count, 0.9 s horizon, 2 ms step | 30 | 450 | 3 | 40,500 | 4.1 × 10⁷ |
DESIGN arithmetic on those FACTs:
- The quadruped configuration runs in real time on one RTX 4090 in its publication, so an episode of 20 s costs about 0.33 GPU-minutes with one design occupying the whole card. The humanoid configuration at the same step rate costs about 1 GPU-minute. At a 4 ms planning step the figures are about 1.7 and 5.2 GPU-minutes.
- A policy rollout costs 500 world-steps per simulated second at a 2 ms step. The sampling controller costs N_s·N_it·(T_h/h_p)·50 planning steps for the same second: about 4 × 10³ to 2 × 10⁴ times as many steps for the rows above, and about 2 × 10⁵ if the 2,048-sample humanoid row plans at the rollout's own 2 ms step. One device at the 2 M world-steps/s target (§36) drives about one world in real time on the cheapest row.
- Training once costs under 10 GPU-minutes for the quadruped and under 60 for the humanoid, after which an episode costs about 0.01 GPU-seconds in a batch. The break-even is therefore about 30 episodes per quadruped design and about 60 per humanoid design, and 6 and 12 at a 4 ms planning step.
- Acceptance needs many more episodes than that. Bounding a failure rate at 1% with zero failures at 95% confidence takes ⌈ln 0.05 / ln 0.99⌉ = 299 episodes: about 99 and 299 GPU-minutes under sampling control, against about 10 and 60 to train once. As an evaluator it is about ten and five times more expensive.
- As a screen it is cheap: one episode costs 30 to 60 times less than training, three episodes 10 to 20 times.
Rules that follow.
- A screening rung of at most three episodes per design by default. A paired screening comparison may use up
to the break-even count of its class (30 for the quadruped setting and 60 for the humanoid at a 20 ms planning
step; 6 and 12 at 4 ms), and stays labeled
screening: sampling-control. - Never a stand-in evaluator. Episodes whose controller kind is
mpc-sampleare refused as the scoring source of acceptance, promotion or qualification (FROB-REF-817, INV-REFLEX-03), unless the task declares the sampling controller as the delivered controller and the episodes meet the four conditions of the delivered-controller lane (§22.22). - Plant equals model. The planner holds the model it is controlling. Its success says what the body can do
under a clairvoyant model, not what a deployable policy will do. The result carries the flag
model-privilegedunless the planning model's parameters were drawn independently of the plant's randomization and the planner ran on estimated state. - Declared latency. The planner's compute time is a modeled delay of one control period unless a measured figure for a named computer is supplied. A zero-delay planner is labeled as such.
The two roles.
- Out-of-distribution fallback. When the baseline is
not-synthesizableorsynthesized-failed, and the family policy of §23.11 reports the design out of distribution, the sampling controller gives a first behaviour to look at, with telemetry, in one to three episodes. - Demonstration source. Its rollouts are demonstrations for distillation (§23.10). The arithmetic limits this too: at the cheapest row and the throughput target, 4,000 simulated seconds of demonstrations cost about 67 GPU-minutes, which is more than one quadruped training run. It pays when pooled: 200 designs at 20 s each for a family prior, against 200 specialists at about 7 minutes each (about 23 GPU-hours), or for a class that has a morphology-independent cost and no shaped reward yet.
Verb: control.mpc (run, with --episodes capped at three unless --demonstrations names a distillation job,
--compare names a paired screening comparison, or --freeze writes a planner record).
Gate: GATE-REFLEX-06. Kill gate: KILL-MPC.
22.9 Gaits as orbits (frob-gait) F
Analysis A-04 listed the gaits a body can have. frob-gait computes whether a listed gait exists as a periodic
orbit of the closed loop, whether it is stable, how it moves when the design changes, and what a design change
does to a metric measured on it.
A screening signal, with its published cost. FACT (Appendix H): published full-order orbit solves take about
4 to 260 s of CPU at 20 to 68 states with initial guesses supplied, which is about 1% to 63% of a 410 s incumbent
quadruped training run. Continuation is not guaranteed to arrive, because turning points and bifurcations stop
it. An orbit is not an accepted controller: published hardware gaits needed added heuristics to stabilize.
Every result of this section is therefore labeled screening: gait-orbit. Floquet multipliers describe the
chosen feedback law on the chosen tier, and local stability only. A locally stable orbit on flat ground supplies
no estimate of recovery on a staircase.
22.9.1 The map
The state is x = (q, v, a, c): configuration, velocity, actuator and thermal states, and controller state (the gait clock, and the filter state when the loop runs on estimated state). One control period on a batched tier is a map F_θ: x ↦ x⁺, with θ the design parameters. The baseline's gait clock forces the loop periodically, so a stride of N control periods is the stroboscopic map P_θ = F_θ^N, and the stride period is a rational number in the clock contract of §20.
On flat ground under a body-frame velocity command the closed loop is equivariant under horizontal translation, and under yaw when heading is not tracked. Those coordinates are removed and listed in the result; x̄ is the state without them. A gait is a fixed point P̄_θ(x̄*) = x̄*, which is a relative periodic orbit of the full state whose drift is the stride displacement.
22.9.2 Shooting with symmetry, stability without it
For a gait with spatio-temporal group H and spatial group K, H/K is cyclic of some order m with generator σ (C-10), and the orbit satisfies the twisted condition F̄^(N/m)(x̄*) = ρ(σ)·x̄*. Shooting solves r(x̄) = ρ(σ)⁻¹·F̄^(N/m)(x̄) − x̄ = 0 for x̄ in Fix(K) by Newton with a line search, seeded by the atlas's phase lags and a settled stance. The problem is m times shorter and of lower dimension than the full stride.
Rule: stability is computed over the full reduced state, never over Fix(K) alone. The monodromy matrix is
M = D P̄_θ(x̄*) = (ρ(σ)⁻¹·D F̄^(N/m)(x̄*))^m with the Jacobian taken in every direction (PO-GAIT-01). M is
block-diagonalized by the isotypic projectors of C-09, so each multiplier is attributed to a symmetry type, and
a multiplier outside the unit circle in a non-trivial block is a symmetry-breaking instability named by its
block. The spectrum goes to FrankenSim's Floquet analysis through the airlock, which takes a supplied monodromy
and also reports non-normal transient growth (FACT: frankensim/crates/fs-spectral/src/floquet.rs,
analyze_monodromy).
Counterexample, the smallest. Take ẋ = −x and ẏ = +y with the symmetry y ↦ −y. On Fix(K) = {y = 0} every orbit decays with multiplier e^(−T). The full-space multiplier in the antisymmetric block is e^(+T) > 1. A restricted computation reports a stable gait for a robot that falls over sideways.
22.9.3 Derivatives, by tier
- Compliant tiers. F is a composition of kernels, each piecewise smooth. The Jacobian is built column by
column, and each column is a tangent world rolled for one stride beside the nominal one: n tangent worlds for
a reduced state of dimension n (DESIGN estimate: n is 40 to about 130 for the robots in scope). The first
implementation uses central differences in
f64onbatch-cpuat two step sizes with a Richardson check. The second uses the derivative programs of C-14 and is differentially tested against the first. - Referee. On the event-located flow, a perturbation is carried through each transverse event by the saltation matrix Ξ = D R + (f⁺ − D R·f⁻)·D g / (D g·f⁻), for a guard g(x) = 0, a reset R and the vector fields f⁻ and f⁺ before and after. Each event records its chart, its guard margin, its identity and the denominator D g·f⁻.
- What exists. FACT (
frankensim/crates/fs-orbit,fs-time/src/hybrid_sensitivity.rs,fs-mbd/src): FrankenSim has shooting, pseudo-arclength continuation and monodromy, with a shooting module that disclaims hybrid saltations, and a separate saltation correction. It has no hybrid multiple-shooting solver on articulated models and no derivatives of the articulated dynamics with respect to state or design. The compliant-tier route is therefore ours and comes first; the referee route waits on upstream work item FD-FSIM-20 (Appendix F). - Labels. Every derivative carries one label of the set of D-15 and C-16 (§9, §10); this section defines none of
its own. Its four cases map onto that set: a derivative with no event in the stride is
smooth; an estimate on one side of a guard isone-sided; a set-valued result, which carries several selection Jacobians, isgeneralized; and the derivative of a declared smoothed surrogate issmoothed. Compliant-tier Jacobians with an unchanged active set arediscrete-step, event-located ones on the referee aresaltation, and central differences arefinite-difference. Near grazing, where |D g·f⁻| falls below its threshold, an ordinary derivative isrefused(FROB-REF-431) and the offending state, the perturbation directions and the one-sided estimates are returned.
22.9.4 Simultaneous impacts are handled
Trot, bound and pronk put two or four feet down at one instant by symmetry.
- On the compliant tiers nothing is special. Contact force is a continuous function of state, the stride map is
continuous and piecewise smooth, and M is computed as above. One extra check runs: the touchdown times of each
simultaneous group are split by ±δ and the one-sided Jacobians are compared. If they differ beyond tolerance
the result carries both as selection Jacobians and the label
generalized. - On the referee an exactly simultaneous group is one coupled impact solve per island (§16). Nearby trajectories reach the same feet in some order, so the stride map has one smooth piece per ordering. The analysis computes the outcome and the saltation product for every ordering of a group of up to four feet (24 orderings) and for the coupled solve, and reports the order sensitivity: the largest difference between outgoing velocities in the mass norm.
- The verdict is
stableonly if all selections agree within tolerance, or if one quadratic norm contracts under every selection (a common Lyapunov matrix from a small semidefinite program in the shared solver). Otherwise it isorder-sensitive, reported with the range of spectral radii and a finite-perturbation recovery test over sampled disturbances, classTested.
Fixture, by hand. A rigid bar of mass m and central radius of gyration k falls flat at speed v onto two supports at distance a either side of its centre, with plastic impacts. Let β = a²/k². An upward impulse p at one support changes that support's velocity by p·(1 + β)/m and the other's by p·(1 − β)/m.
- Coupled solve: impulses m·v/2 at each support stop the bar. It ends at rest.
- One support first, by an arbitrarily small time: the first impulse is m·v/(1 + β) and leaves the other end approaching at 2β·v/(1 + β). The second impulse stops that end and changes the first end's velocity by 2β·(1 − β)·v/(1 + β)². For β < 1 this is positive: the first end lifts off and the bar is left rocking.
- A uniform bar of half-length l on supports at ±l/3 has β = 1/3: the first end leaves at v/4 and 1/16 of the kinetic energy survives, against none in the coupled solve. For β ≥ 1 the first contact stays closed, both impulses are positive, and every ordering gives rest.
The switch is the sign of the off-diagonal contact-space term J₁·M⁻¹·J₂ᵀ = (1 − β)/m. Outcomes are continuous in the ordering when it is zero or negative and jump when it is positive, which agrees with the published kinetic-angle condition for plastic impacts (FACT, Appendix H). The analysis reports that term for every simultaneous group. On a compliant tier the same bar moves from rest to rocking over a time offset of the order of the contact time constant: a steep, finite slope, which the split check finds.
22.9.5 Continuation in design parameters
Pseudo-arclength continuation of r(x̄, θ) = 0 follows a path in θ: one parameter, or the line between two designs joined by a typed edit. A child design warm-starts from its parent's orbit. What happens on the path is a typed outcome, not an error:
| Outcome | Detection | Consequence |
|---|---|---|
arrived |
Residual below tolerance at the target | The orbit and its spectrum at the target design |
fold |
A multiplier of M at +1 in the trivial block, so I − M is singular and the branch turns back | The parameter value of the fold; no orbit of this type beyond it on this branch |
period-doubling |
A multiplier at −1 | The doubled orbit is offered as a new seed |
torus |
A complex pair crossing the unit circle | Reported; quasi-periodic motion is not followed |
symmetry-breaking |
A multiplier at +1 in a non-trivial block | The broken generator; the lower-symmetry branch is offered as a seed |
mode-change |
The orbit's contact sequence changed | The old label no longer applies; a new shooting problem is posed |
diverged, budget |
Newton failure after step halving; the budget ended | Nothing partial is published as an orbit (§35) |
22.9.6 Design gradients
At a fixed point with I − M nonsingular, d x̄*/dθ = (I − M)⁻¹·∂P̄/∂θ. For a metric J(x̄*, θ) accumulated over the stride (cost of transport, peak torque margin, friction utilization), dJ/dθ = ∂J/∂θ + λᵀ·∂P̄/∂θ with (I − M)ᵀ·λ = ∇J. This is the implicit-function theorem at a fixed point. It costs one more tangent world per design parameter and one small solve (PO-GAIT-02). It is refused at a fold (FROB-REF-431). The gradient of a multiplier is taken by differences along the continuation path. Where a gradient is refused, the caller is routed to derivative-free search in §28.
22.9.7 What a result contains
A GaitOrbit record holds the ControllerId, the tier, (H, K, σ), the
stride length N, x̄*, the residual, the removed coordinates, the multipliers with their block labels, transient
growth, order sensitivity, the derivative labels, the metrics, and the standing screening (INV-GAIT-01).
Verbs: gait.atlas (enumeration and phase lags; also analysis A-04), gait.orbit, gait.continue,
gait.gradient. Gates: GATE-GAIT-01 to -03. Kill gate: KILL-GAIT.
22.10 Skill artifacts (.fskill)
What it is. A skill is a behaviour stated at the task boundary, independent of the body that performs it.
Format FMT-FSKILL, extension .fskill, schema frob.skill/1 (DESIGN sketch; values illustrative):
{
"schema": "frob.skill/1",
"name": "push-cart-level",
"task_frame": "cart_handle",
"interface": { "command": "wrench", "observables": ["handle_pose", "handle_twist"], "clock_hz": "50/1" },
"phases": [
{ "id": "approach", "exit": "contact(handle)", "reference": "twist-profile#a1" },
{ "id": "push", "exit": "distance_to_goal < 0.15 m", "reference": "wrench-profile#p1" }
],
"assumptions": { "mu_min": 0.4, "payload_kg": [0, 40], "slope_max": 0.05 },
"secondary_costs": ["electrical_power", "winding_temperature"],
"exact_layer": { "map": "constrained-mass-solve", "allocation": "cone-program", "witness": "dual" },
"residual": null,
"evidence": { "class": "Tested", "authorization": "the no-authorization flags of §12" }
}
Layer 1, exact S. For a morphology and a contact mode m, the equations of motion M·v̇ + h = Sᵀ·u + J_cᵀ·f_c with the mode's contact constraints give, after a constrained mass solve that eliminates v̇ and f_c, an affine map from actuator effort to the task-space quantity: w = w₀(q, v) + B_m(q)·u. The allocation program minimizes ‖B_m·u − (w* − w₀)‖² plus the secondary costs, subject to the actuator boxes, the contact cones (with μ_min for forces the robot commands) and unilateral support limits. It is a cone program in the shared solver and it returns one of two things:
- a feasible-allocation witness (u, f_c), which replays; or
- before any training, an infeasibility witness: a direction c in task space, the support value max cᵀ(w₀ + B_m·u) over admissible u, which is below cᵀw*, the gap, and the binding actuator or contact (FROB-REF-448). This is the dual certificate of §22.5, checked exactly.
Smallest case, by hand. A planar two-link arm has link lengths 0.4 m and 0.3 m and joint angles 0 and π/2. Its tool point is (0.4, 0.3) m and its Jacobian is J = [[−0.3, −0.3], [0.4, 0.0]] m/rad, so a tool force F maps to joint torques τ = JᵀF = (−0.3·F_x + 0.4·F_y, −0.3·F_x). A tool force of (10, 0) N needs torques (−3, −3) N·m. If each actuator is limited to 2.5 N·m, the second row alone gives |F_x| ≤ 2.5/0.3 = 25/3 N. The witness is the direction c = (1, 0), the support value 25/3 N, the requirement 10 N and the binding actuator, joint 2. Taking the lower rows of a space-frame screw Jacobian in place of the point Jacobian fails this fixture.
Limits written into every artifact.
- Instantaneous. A feasible witness at each sampled state of a reference is necessary and not sufficient. The map cannot show multi-step dynamic feasibility such as capture-point recovery or a momentum-assisted jump.
- Scope. Task-space skills of wheeled, wheel-legged and manipulating robots, and the stance phases of legged ones. Legged locomotion needs a phase sequence to transfer, which §22.9 and §23.12 address.
- Floating base. A fixed-base pullback τ = JᵀF is wrong for a floating-base robot. The map must come from the contact-consistent constrained solve.
- Refinement. A transported skill refines its specification only if, for every admitted target state, every abstract command and every target disturbance, each target successor is related to an allowed abstract successor, with the command chosen from information available before the disturbance. Counterexample: a target whose successors are {stable, fall} cannot refine an abstract transition whose only successor is stable by selecting its stable branch (FROB-REF-444).
- On failure the report returns the contact mode, the requested wrench, the support function of the attainable set in the failing direction, and a counterexample that re-runs.
Layer 2, learned residual transport M. A frozen, morphology-conditioned residual acts in task space above the allocation program. Its inputs are made dimensionless by the body's mass m₀, a length ℓ₀ and gravity g, with momentum scale m₀·√(g·ℓ₀) and angular-momentum scale m₀·ℓ₀·√(g·ℓ₀). Dimensionless inputs give dynamic similarity only when every dimensionless group matches; published similarity transfer shows no loss on dynamically similar contexts for a pendulum and a planar runner only (FACT, Appendix H). The experiment has three treatments at equal compute and equal source data: training from scratch on the target, a morphology-conditioned policy with full transfer learning, and interface-based transfer with limited adaptation. Both denominators are reported: the incremental cost per target and the amortized cost with pretraining. It is off the critical path; its pilot, ceiling and kill gate are in §22.17.
Verbs: skill.compile (a .fskill and a design give the per-mode maps and the program), skill.check (a
witness of either kind over a set of states), skill.transport, skill.show. Gates: GATE-REFLEX-08, -09.
skill.transport moves a skill artifact to another body with the learned residual of layer 2 and is M.
mission.transport (§25.3) is a different verb: it moves a certified claim through an interface map with a bounded
error and is F. Neither stands in for the other.
22.11 The energy-budget safety tier (.fenergy) F
The idea. Certifying a contact sequence is hard. Bounding the energy that can reach a boundary is not. Passive contact, with any friction coefficient μ ≥ 0, any restitution between 0 and 1 and any sequence of impacts, can dissipate or redistribute mechanical energy and cannot create it. So if everything that could deliver energy to a boundary is bounded, and the absorber at that boundary can take that much, the stop holds for every contact sequence. No contact model appears in the argument.
The budget. For a declared boundary (a stop, a bumper, a door or drawer at the end of its travel, a lift platform, a compliant cover that may touch a person) and a time t:
E_deliv(t) = T_ub(t) + E_drive(t) + E_ext + E_stored(t) + E_reserve ≤ W_abs
| Term | Definition | Source |
|---|---|---|
| T_ub | Upper bound on the kinetic energy of every body on the load path, ½·v̂ᵀ·M̄·v̂, with measured speeds inflated by their sensing error and masses at their caps. It includes reflected rotor inertia. | Model, encoder cards |
| E_drive | ∫ max(P_elec, 0) dτ over the cutoff delay Δt_cut, bounded by P_max·Δt_cut: what the drives can still inject before the cutoff acts. Δt_cut is detection plus transport plus driver plus brake delay. | Actuator cards; clock contract of §20 |
| E_ext | Upper bound on work done by sources that are not cut: gravity over the stroke, a conveyor, another machine | Scene and declaration |
| E_stored | Elastic and other stored energy that can be released onto the boundary: series springs, preloads | Model |
| E_reserve | Numerical and sensing reserve, at least the energy ledger's residual bound over the window | Ledger of §29 |
| W_abs | Lower envelope of the work the absorber takes before its limit (bottoming, or a stated force limit) | Interface card with provenance |
The monitor. One accumulator and one comparison per control tick: kernel K-energy-monitor, in the
exact-integer kernel class of §35 with declared unit scales, so the admission predicate is a strict inequality
evaluated exactly on every backend. It outputs the margin W_abs − E_deliv. At zero margin it cuts drive power
and hands over to the backup controller of §22.12. The longest admissible cutoff delay is a closed form:
Δt_cut,max = (W_abs − T_ub − E_ext − E_stored − E_reserve) / P_max.
Premises, each checked or declared.
- Sources are complete. Every port of the design's port graph (C-12) that can inject power into the load
path is in the budget with a bound, or is cut by the monitor. The check is mechanical:
shield.budgetwalks the port graph and refuses a budget with an unlisted source (FROB-REF-451). - Caps are card values. Mass caps and speed sensing errors come from cards with provenance.
- The cutoff acts within Δt_cut. If the packet-delay distribution has a tail beyond Δt_cut, the claim becomes a probability and the artifact says so.
- The absorber is on every forward path to the boundary. This is a geometric premise, taken from a declared guide or from the clearance analysis of §19. Without it the claim covers only contacts made through the absorber.
What the claim is. "For every contact sequence and friction coefficient, the energy delivered to the
declared boundary during a stop does not exceed the absorber's stated minimum capacity, in the model, under
premises 1 to 4." The inequality is an exact-integer predicate that the offline auditor recomputes
(PO-SHIELD-01). Its class follows the claim product of §8: Certified for the arithmetic and no better than the
weakest input, so an absorber envelope marked assumed yields an Estimated claim. Reality is no-claim until
bench data for the absorber and the drive are imported through Ground (§30). One absorber measurement and one
drive measurement would then serve every later policy for that hardware, since no term depends on the policy.
What it does not support.
- Peak force, pressure or injury thresholds, except at the absorber and only if W_abs was defined up to a force limit.
- Quasi-static clamping after the stop. A drive that keeps pushing poses a force question; the budget answers an energy question.
- Contacts that bypass the absorber, and sources that are not in the port graph: a second robot, a falling object, a person pushing.
- Regulatory conformity, or permission to operate near people. The artifact carries the no-authorization flags of §12. FACT (Appendix H): a published technical specification for collaborative robots states transient contact limits through transferred energy. The budget is a quantity of that kind; it is not an assessment against any such document.
- Anything about the simulator's contact law. The simulator is used the other way round, as a negative control that must never show an exceedance for any μ, any restitution or any contact law.
Worked card (illustrative values; the arithmetic is the content). A cart of 8 kg effective mass approaches a stop at 0.4 m/s: T_ub = ½·8·0.4² = 0.64 J. The drive can inject 20 W during a 25 ms cutoff delay: 0.50 J. The floor slopes by at most 8% over a 40 mm absorber stroke: 8·9.81·0.040·0.08 = 0.25 J. A series spring holds 0.10 J. The total is 1.49 J against an absorber rated at 50 N over 40 mm, W_abs = 2.00 J. The longest admissible cutoff delay is (2.00 − 0.99)/20 = 50.5 ms, or 45.5 ms with a further 0.10 J of reserve.
Counterexample to the obvious shortcut. Budgeting the payload's kinetic energy alone omits the rotors. A rotor
of inertia 5 × 10⁻⁵ kg·m² behind a 50:1 reduction, driving a wheel of radius 0.05 m, has an effective mass of
50²·(5 × 10⁻⁵)/0.05² = 50 kg, more than six times the 8 kg cart. shield.budget computes T_ub from the full mass
matrix, in which reflected inertia already appears (§16).
Format FMT-FENERGY, extension .fenergy, schema frob.energy/1 (DESIGN sketch; quantities are decimal strings
so that no binary rounding enters the predicate):
{
"schema": "frob.energy/1",
"boundary": "dock_stop_A",
"load_path": ["chassis", "payload", "drive_left", "drive_right"],
"unit_scale": { "energy": "1e-6 J", "time": "1e-6 s" },
"terms": {
"kinetic_ub": { "mass_cap_kg": "8", "speed_error_mps": "0.01", "includes_reflected_inertia": true },
"drive": { "p_max_w": "20", "cutoff_delay_s": "0.025" },
"external": [{ "source": "gravity", "bound_j": "0.25", "basis": "slope <= 0.08 over 0.040 m" }],
"stored": [{ "source": "series_spring", "bound_j": "0.10" }],
"reserve_j": "0.00"
},
"absorber": { "card": "bumper-50N-40mm", "w_abs_j": "2.00", "provenance": "assumed" },
"premises": { "sources_complete": "port-graph-digest", "absorber_on_path": "declared-guide" },
"claim": { "class": "Estimated", "reality": "no-claim", "authorization": "the no-authorization flags of §12" }
}
The sketch is the worked card with no reserve: kinetic 0.64 J, drive 0.50 J, external 0.25 J and stored 0.10 J sum
to 1.49 J against 2.00 J. Declaring reserve_j of 0.10 raises the left side to 1.59 J and shortens the longest
admissible cutoff delay from 50.5 ms to 45.5 ms.
Verbs: shield.budget (build and admit a budget for a design and a boundary), shield.check (replay a run
against a budget). Gate: GATE-SHIELD-01. Kill gate: KILL-ENERGY.
22.12 Runtime shields (frob-shield) F
What a shield is. An ActionFilter: a stage between the policy or the baseline and the actuator limiter,
which replaces the proposed command u_π by the nearest command that keeps a set of barrier rows satisfied. It is
a runtime filter. Its results are class Tested, typed as a claim product, and nothing converts them into a
proof (INV-SHIELD-02).
22.12.1 The filter
Minimize ‖u − u_π‖² in a metric R, subject to A(x̂)·u ≥ b(x̂) + δ and lo ≤ u ≤ hi. Each row is the condition ḣᵢ + α(hᵢ) ≥ 0 for a barrier function hᵢ ≥ 0 on a reduced-order model (22.12.4), written linearly in u. The margin δ has three declared parts: an inter-sample term L·Δt, where L bounds the time derivative of the row's left side over a control period; an estimation-error term; and the measured template-to-full term of 22.12.4. Two rules of formulation:
- Quantifier order. There exists one command, chosen from delivered information, that satisfies the rows for every admitted disturbance and estimation error. The reverse order lets the command depend on the disturbance, which no controller can do.
- Split cones. Contact forces the robot commands are limited to the inner cone K(μ_min). Forces the environment may apply range over the outer cone K(μ_max). Taking the infimum over the inner cone for forces the environment chooses is optimistic.
22.12.2 The projection
- One row, diagonal metric: exact. u(λ) = clip(u_π + λ·R⁻¹·a, lo, hi), and aᵀu(λ) is nondecreasing and piecewise linear in λ with at most 2n breakpoints. Sort them and take the smallest λ ≥ 0 with aᵀu(λ) ≥ b + δ (PO-SHIELD-02). If the maximum of aᵀu over the box is below b + δ, the row cannot be met within the actuators and the backup runs.
- Counterexample to the closed form that ignores the box. R = I, u_π = (1, 0), box [0, 1]², row u₁ + u₂ ≥ 1.5. The unboxed projection is (1.25, 0.25), which leaves the box. Clipping it gives (1, 0.25), which violates the row. The breakpoint solution is λ = 0.5 and u = (1, 0.5).
- Two to four rows. A fixed number of iterations of the shared solver, then a mandatory optimality check:
primal feasibility of rows and box, nonnegative multipliers, complementarity and stationarity, each within a
stated tolerance. If any part fails, or the rows conflict, the step invokes the backup controller atomically.
A partially converged iterate never reaches a driver: only a
Checked<Command>can be passed on (PO-SHIELD-03, INV-SHIELD-01). - More than four coupled rows are refused when the shield is fitted (FROB-REF-454).
22.12.3 Backup controllers and monitors
Backups live in frob-control, each with a declared domain.
| Class | Backup | Property |
|---|---|---|
| Wheeled, tracked | Damping injection on the drive outputs, then brakes where the card has them | With collocated outputs y and K_d ⪰ 0, the continuous-time command u = −K_d·y gives dH/dt ≤ −yᵀ·K_d·y ≤ 0 in the port model of C-12. A command sampled and held for a period h keeps the stored energy from increasing at the ticks only under the gain bound 2·M̄ − h·K_d ⪰ 0 on the reflected inertia M̄ (PO-SHIELD-04, and the sampled-damping subsection of this chapter): 300 N·s/m held for 10 ms on a free 1 kg slider turns +1 m/s into −2 m/s and adds 1.5 J. Holding on a slope needs the brake's holding torque, and the domain says so. |
| Arm | Posture hold by energy shaping, with damping | Passive with respect to the shaped energy, for a fixed base |
| Legged | Widen the stance, lower the body, freeze | Tested only; no passivity claim once contacts switch |
Monitors watch the assumptions: a friction estimate outside [μ_min, μ_max], innovation spikes in the estimator, unmodeled impact impulses, tracking error above the calibrated margin, actuator saturation. An alarm also triggers the backup. The violation bound per episode is
P(violation) ≤ E[N_breach]·δ₁ + P(N_alarm ≥ 1)·δ₂ + δ₃
with δ₁ = E[N_silent]/E[N_breach], the expected number of breaches with no alarm over the expected number of breaches (zero when no breach is expected), δ₂ = P(violation in the episode | N_alarm ≥ 1), which covers several alarms in one episode and a backup that fails late, and δ₃ = P(violation with no breach and no alarm). Each is estimated with an interval on independent hidden-state labels from the simulator, never from the monitor's own output. Counterexample to self-labeling: a monitor that is silent on every breach it cannot see reports δ₁ = 0 from its own log. The first term is Markov's inequality on the number of silent breaches and the second is a conditional probability, so no term grows with the number of alarms in an episode. The three numbers are estimated for one deployed policy and are estimated again when the policy changes. The bound is not claimed under a shift of the world distribution, because a shift changes δ₁ itself; shifted worlds are reported as a separate row.
22.12.4 Reduced-order barrier filters with a measured margin
Barriers are written on a template of 2 to 12 states: a single integrator or a unicycle for a base (keep-out zones, speed limits, door-pinch zones), a linear inverted pendulum or a capture point for a legged body. A tracking controller realizes the template's filtered command on the full robot. The template's guarantee reaches the robot only through the tracking error ε between commanded and realized template state, and ε depends on inputs, disturbances and the full state, so it cannot be computed in advance (FACT, Appendix H). It is therefore measured, and labeled as measured:
- Calibration. n episodes on a fast tier across the calibrated family. The score of an episode is its largest tracking error.
- Margin. ε̂ is the ⌈(n + 1)·(1 − α)⌉-th smallest score. For an episode exchangeable with the calibration set, the error is at most ε̂ with probability at least 1 − α. The barriers are tightened by ε̂.
- Scope. The margin is keyed by design class, controller or policy identity, task family and tier. A new policy, or a design outside the class, has no margin, and the shield refuses to run on a borrowed one (FROB-REF-457).
- Drift. A running test on the conformal p-values raises an alarm when exchangeability fails. The regime ends and calibration restarts inside the new one.
- Tier. A sample of calibration episodes is replayed on the referee, and the tier difference from the ledger of §16 is added to ε̂.
FACT (Appendix H): reduced-order barrier filters run on real legged and aerial hardware under a tracking
assumption, and their published hardware results say constraints are "rarely violated". They are strongest for
wheeled and tracked bases and for slow legged motion. No full-order certificate is attempted here:
sum-of-squares methods reach about five to ten states; a diagonally dominant relaxation reached 30 states only
for balancing with pinned feet and no input limits; and FrankenSim's certify_roa uses a fixed quadratic
Lyapunov function with a largest example of two states (FACT: frankensim/crates/fs-sos/src/roa.rs).
22.12.5 Reachable-set tables for small subsystems M
For a subsystem of at most four reduced states with
a bounded sensing error E (a wheel at a threshold, a tray about to tip, braking across two floor materials), a
table over at most 8 modes × 16,384 cells is built by backward viability pruning. A cell K_m keeps a common safe
action only if R_mn((K_m ⊕ E) ∩ G_mn, D) ⊆ K_n ⊖ (−E) for every guard G_mn that may be enabled (cells hold measured states, the true state is the measured state plus an error in E, and A ⊖ B = {x : x + B ⊆ A}; the erosion is by −E, which equals E only for a centrally symmetric error set; where cells hold true states and the whole belief set must stay in one cell, the condition is R_mn(K_m ∩ G_mn, D) ⊆ K_n ⊖ (E ⊕ (−E))), where R_mn is the
reset, D the disturbance set, and inter-sample sweep, latency and chatter are accounted for. At run time the
whole belief set is looked up; outside the table the answer is uncovered and the backup domain applies. The
reduction of a robot to four states is an assumption that the monitors of 22.12.3 watch. It is not proved. The
table is a payload of the shield record, with no file format of its own. It is off the critical path.
22.12.6 Shields during training
The filter is a stage of the environment. The learner's sampled action is
the command before the shield, its likelihood is the likelihood of that command (§23.6), the applied command is
logged beside it, and each intervention is an event that a reward term may price. A shield that intervenes on
more than a declared fraction of ticks sets the flag shield-dominated.
A rule for every adaptive element here (the margin calibrator, the drift alarm, the alarm thresholds): it ships with a replay scenario, evidence records, a calibration report and a demonstrated fallback, or it does not ship.
Verbs: shield.fit (rows, margins, backup and monitors for a design, a controller and a task family),
shield.check, shield.report. TARGETs: at most 5% added step cost; at least 85% of unshielded return kept.
Gates: GATE-SHIELD-02 to -05. Kill gates: KILL-SHIELD, KILL-TUBE.
22.13 Symmetric controllers under stabilizer intersection S
Two groups. Γ_mech is the group of C-09: automorphisms of the design that preserve link lengths, pseudo-inertias, joint screws with their reflection parity, actuators and gear ratios. Γ_obs ≤ Γ_mech is the subgroup that also preserves every sensor pose and field of view, the payload, the task's goal, and the symmetry of terrain and guards. Both are stored on each robot-plus-task pair. A quadruped built symmetric fore-aft and left-right has Γ_mech = C₂ × C₂; with a front-only lidar its closed loop has Γ_obs = C₂.
The group a calculation may use is the intersection of Γ_obs with the stabilizers of what that calculation holds fixed, computed when it is used:
Γ_use = Γ_obs ∩ Stab(gravity in the body frame) ∩ Stab(boundary conditions)
∩ Stab(current contact mode) ∩ Stab(commanded twist)
On a side slope the left-right mirror leaves Stab(gravity) and the controller's left and right gains untie. With an off-centre payload the mirror is not in Γ_obs at all. A reduction requested under a larger group is refused (FROB-REF-461, INV-REFLEX-04), and the full model always stays executable.
Rules.
- Parity. Under an orthogonal matrix P a polar vector maps by P and an axial vector by det(P)·P. A mirrored leg may need a sign change at its encoder. Reflections are not elements of SE(3); they act on the design term, not through a frame transform.
- Exact reduction against distribution symmetry. Tying gains or block-diagonalizing a linearization inside one world needs that world and its data to be invariant. A symmetry of the distribution of worlds pairs two different worlds: it supports augmentation and paired tests (§23.13) and no reduction inside either world.
- Linear blocks only. A linearized operator that commutes with the action preserves the isotypic blocks. An equivariant nonlinear map need not decouple into independent blocks.
- Equivariance in law. Fixture: two identical latches, exactly one must close, and the observation is fixed by the mirror that exchanges them. A deterministic equivariant controller must issue matching commands, and both matching commands fail. A controller that draws one of the two successful choices uniformly is equivariant in law. A tied deterministic controller that fails here has excluded only its own class. The controller's manifest states whether pathwise or in-law equivariance is required and who shares the draw.
- Asymmetric worlds need asymmetric actions. A quadruped crossing a patch that is slippery on one side must be free to act asymmetrically. Terrain is part of the observation, so an equivariant controller is conditioned on the transformed terrain; it is not constrained to mirror-symmetric commands.
- Symmetry at an event. A symmetry saves work across an impact only if it commutes with the event. For a group element g and a transverse event e with saltation Ξ_e the identity is D g(x⁺)·Ξ_e(x⁻) = Ξ_(g·e)(g·x⁻)·D g(x⁻). Test: reflect a four-foot landing and exchange the left and right sole laws; a one-sided perturbation transformed before or after the impact must predict the same change. The identity transports a local derivative to the corresponding event. It does not license block reduction of a trajectory that the symmetry does not fix.
- A change of group at a mode switch. Γ_use is recomputed when the contact mode changes, so gains tied in one mode untie in the next. Nothing here requires the command to be continuous across that change (DESIGN). An exact reduction is the full-model calculation carried out under a group that leaves it invariant, so in each mode the reduced controller issues the commands the full model issues, and whatever jump the command makes at the switch is the jump of the mode-scheduled law itself, the same with the reduction on or off. The check is that comparison on both sides of the switch, within the determinism class; a disagreement means a wrong stabilizer and is refused (FROB-REF-461). A law that needs a continuous command across modes states that as its own requirement; symmetry neither supplies it nor removes it.
Verb: control.symmetry returns Γ_mech, Γ_obs, the stabilizers in force and Γ_use, with the generators each
stabilizer removed. Gate: GATE-REFLEX-07.
22.14 Refusals
Each refusal has a stable code, a message and ranked repairs (§35). Reflex uses the 4xx family, and one code
in the evidence family.
| Code | Refused | Ranked repairs |
|---|---|---|
| FROB-REF-421 | No synthesis route exists for the detected morphology class and no fallback is admitted | Add the ports or sensors the class requires; admit the sampling controller for screening; go to rung R4 |
| FROB-REF-424 | A control program is infeasible and the controller has no admissible fallback | Read the dual certificate; relax the task; change the binding actuator or contact |
| FROB-REF-431 | A gait derivative was requested at grazing or at a fold | Use the one-sided estimates returned; move the parameter; derivative-free search |
| FROB-REF-448 | A skill's task wrench lies outside the body's attainable set | The witness names the binding actuator or contact: regear, reshape, or change the wrench profile |
| FROB-REF-444 | A transported skill does not refine its specification | The counterexample names the uncovered successor: add a guard or narrow the entry set |
| FROB-REF-451 | An energy budget omits a source in the port graph, or a class above Estimated is requested on an assumed absorber envelope |
Bound the source or cut it; import the absorber measurement |
| FROB-REF-454 | A shield's rows have no common admissible command on the calibration set, or more than four are coupled | Reduce the row set; widen the actuator box; lower the speed limit |
| FROB-REF-457 | A shield margin is used outside its scope key | Recalibrate for the new policy, class, task family or tier |
| FROB-REF-461 | A symmetric reduction was requested under a group not contained in Γ_use | Use Γ_use; remove the asymmetry with the typed edit that A-03 proposes |
| FROB-REF-464 | An estimator was scored with a fitted scale, or an estimated-state run lacks a sensor the filter needs | Score with translation and yaw alignment only; add the sensor, or run privileged and say so |
| FROB-REF-817 | Sampling-controller episodes were offered as the scoring source of acceptance, promotion or qualification | Train or fine-tune a policy for the finalist; keep the sampling result as screening |
22.15 Invariants and proof obligations
| ID | Invariant |
|---|---|
| INV-REFLEX-01 | Every command that reaches a driver records the layer that produced it, and every fallback is a telemetry event with its reason. |
| INV-REFLEX-02 | A ControllerId covers everything that affects the controller's output: class, gains, programs, iteration counts, planning configuration, latency model and the group used. |
| INV-REFLEX-03 | Sampling-controller results never enter acceptance, promotion or qualification evidence. |
| INV-REFLEX-04 | A symmetric reduction uses only the stabilizer intersection stored for the robot-plus-task pair. |
| INV-GAIT-01 | Every orbit result carries its ControllerId, tier, residual, full-space spectrum, derivative labels and the standing screening. |
| INV-SHIELD-01 | A filtered command satisfies the actuator box exactly and the active rows within tolerance, or the backup controller is running. |
| INV-SHIELD-02 | A shield result is a claim product of class Tested; no conversion to Proved or Certified exists in the type. |
| ID | Obligation | Verification |
|---|---|---|
| PO-REFLEX-01 | The filter's linearization has the four-dimensional null space of position and yaw at every estimate. | Rank test on 10³ trajectories; finite perturbation of the true initial state |
| PO-REFLEX-02 | A zero residual reproduces the baseline's digests in class D2. | Bit comparison on the smoke battery |
| PO-REFLEX-03 | Every infeasibility certificate checks in exact rational arithmetic: Aᵀy ∈ K* and bᵀy < 0. | The offline auditor of §12 re-checks each emitted certificate |
| PO-GAIT-01 | The monodromy assembled from the twisted factor equals the direct Jacobian of the full stride. | Both computed in f64; relative difference below 10⁻⁸ |
| PO-GAIT-02 | Design gradients agree with central differences away from labeled non-smooth points. | The 100-case suite of GATE-GAIT-02 |
| PO-SHIELD-01 | The energy inequality is an exact-integer predicate with the same value on every backend. | Recomputed offline from the .fenergy file and the run's integer telemetry |
| PO-SHIELD-02 | The single-row projection equals the exact minimizer and never leaves the box. | Rational reference on 10⁶ instances |
| PO-SHIELD-03 | No command leaves the filter without passing the optimality check or coming from the backup. | Type-state in the API; a compile-fail test; a planted non-convergent instance |
| PO-SHIELD-04 | Under damping injection the port model's stored energy is non-increasing: in continuous time for every K_d ⪰ 0, and at the ticks of a sampled, held command when 2·M̄ − h·K_d ⪰ 0 (with actuation latency, when the exact one-tick map of A-22 contracts in the energy norm). | A polynomial identity for Lemma (a theorem card in §12) for the continuous statement; an exact rational check of the gain bound per fitted backup; the ledger's monotonicity on replay |
22.16 Known answers and negative controls
Known answers are numbered in §38. Each row below is hand-checkable and seeds the registries of §38 and Appendix N.
| Fixture | Expected answer | Negative control |
|---|---|---|
| Filter observability | Null space of dimension four | A filter linearized at a moving estimate in the conventional way gains a spurious yaw direction (FACT, Appendix H); the rank test must see it |
| Gauge scoring | A 2 m estimate of a 1 m motion scores 1 m | Scale-fitted scoring reports zero and is refused |
| Friction certificate | y = (0.5, −1), yᵀb = −5 N | A certificate nudged outside the dual cone fails the exact check |
| Two-link arm witness | F_x ≤ 25/3 N from joint 2 | The lower rows of a space-frame Jacobian give a different bound |
| Zero residual | Digests equal the baseline's | A residual with a nonzero bias changes the digest |
| Sampling cost rows | 65,536, 204,800 and 40,500 steps per tick; 299 episodes for 1% at 95% | A planted acceptance request is refused |
| Rimless wheel on a slope γ with spoke half-angle α | With c = cos 2α and b = (4g/l)·sin α·sin γ: ω*² = c²·b/(1 − c²), multiplier c² | An uphill slope (b < 0) has no fixed point |
| Planted transverse instability | Full-space multiplier e^(+T) | The restricted computation says stable and must be marked incomplete |
| Bar on two supports | Rest when coupled; first end leaves at v/4 for β = 1/3; rest for β ≥ 1 | A solver that forces an order reports rocking for the coupled case |
| Energy drop-stop card | 1.49 J ≤ 2.00 J; 50.5 ms; 45.5 ms | A payload-only budget omits 50 kg of reflected mass; an unlisted spring is refused |
| Shield projection | (1, 0.5) | The unboxed closed form gives (1.25, 0.25) |
| Damping-injection backup | Stored energy non-increasing in the ledger at every tick, for gains inside the sampled bound | A negative damping gain breaks monotonicity and is caught; so does a gain above the bound (300 N·s/m held 10 ms on a free 1 kg slider: +1.5 J) |
| Two latches | The in-law controller succeeds; the tied deterministic one fails | — |
| Planted asymmetries | A side slope removes the left-right mirror; a front lidar leaves C₂ | Tying under Γ_mech loses return on the one-sided patch |
| Skill refinement | A {stable, fall} target is refused | — |
22.17 Pilots and ceilings for the Frontier and Moonshot items
Each pilot is paper-scale, runs with existing tools outside the workspace as a development-time experiment (§38), and must run before its work package is funded. The ceiling is stated before any speed test.
| Item | Tier | Pilot | Falsifier | Ceiling of the benefit | Cost (DESIGN estimate) |
|---|---|---|---|---|---|
| Frontier rows of §22.4 (predictive, biped balance, hybrid, arm on a base) | F | Each controller on one reference robot of its class on the referee, on that class's first task of §24 | The class fails its row of GATE-REFLEX-01 after one repair cycle | One more morphology class moves without training; no speed claim | 2 engineer-days per class |
| Baseline plus residual | F | One quadruped with a hand-written trot controller on an incumbent simulator: residual against end-to-end at equal steps, three seeds | The residual needs more than 50% of the specialist's steps for the same return | Training steps per design fall by at most the inverse of the residual's step fraction, for classes that have a baseline | About 10 GPU-hours |
| Sampling controller | F | Thirty designs (two legged templates with limb lengths ±20% and masses ±30%, plus five wheeled), one fixed cost per class, three episodes each on an existing sampling controller; specialists for ten | Success under 60%; a median above 1 GPU-minute (quadruped) or 6 (humanoid) per episode; Kendall τ against specialists below 0.5 | One screening episode costs 30 to 60 times less than training once; nothing as a stand-in evaluator | 2 engineer-days, about 15 GPU-hours |
| Gait orbits and continuation | F | One 12-DoF quadruped trot in existing tools: 20 random 10% perturbations of link lengths and masses, singly and jointly, warm-started from the parent; specialists for ten | Continuation succeeds on fewer than half; median solve above 10 minutes; Kendall τ on cost of transport below 0.5 | One orbit in place of one training run, for periodic gaits on regular ground: 1.6 to about 100 times per design point on published costs (410 s against 260 s and 4 s) | 3 engineer-days, about 4 GPU-hours |
| Energy budget | F | The worked card, and a one-degree-of-freedom cart on the referee with μ, restitution and contact law swept | Any exceedance under the premises | No speed benefit: one class of safety evidence stops depending on the policy | 1 engineer-day |
| Reduced-order barrier shield | F | A 12-DoF quadruped with a trained policy behind a velocity-level filter on a single-integrator base model, 10⁴ randomized episodes | The margin implied by the 99.9th-percentile tracking error exceeds 0.3 m, or violations exceed 10⁻³ per episode | Evidence quality only; no source shows a shorter path to a decision | 2 engineer-days, under 5 GPU-hours |
| Skill transport by learned residual | M | 24 target bodies of one class, three treatments at equal compute and source data | Fewer than 18 of 20 feasible bodies within 3 points of tuned baselines, or one infeasible body not refused | Transfer cost per target against the fastest baseline's; a pretraining cost P is repaid after P/(c_b − c_t) targets | About 40 GPU-hours |
| Reachable-set tables | M | Braking across two floor materials: two states and a mode | Coverage under 90% of the operating set at 16,384 cells per mode, or one false admission in 5,000 in-envelope episodes | A lookup replaces an online reachable-set computation; no design decision is known to depend on it | 3 engineer-days |
22.18 Gates and kill gates
| Gate | Claim | Threshold (TARGET) | At |
|---|---|---|---|
| GATE-REFLEX-01 | Known answer | Each archetype stands for 10 s and traverses 5 m of flat ground under its baseline controller: at M1 the first archetypes (quad-mammal, humanoid, diff-drive and mecanum, §14.10), on referee only, restricted to those whose contact law exists at M1 and whose contact count fits the referee's 21-contact cap (§16.6); at M2 the same set on batch-cpu; at M3 all eighteen on batch-cpu, except tracked and hybrid archetypes, which follow at M4 with their contact laws. |
M1, M2, M3, M4 |
| GATE-REFLEX-02 | Design comparison | BET-06: at least 80% of corpus designs that pass design check traverse 5 m, as a paired 95% lower bound. Potency control: every corpus design with a planted control-relevant defect fails. |
M3 |
| GATE-REFLEX-03 | Known answer | Filter consistency inside χ² bounds on the simulated IMU; position drift under 2% of distance on a flat trot, scored by the gauge rule; PO-REFLEX-01; the comparator's overconfident flag fires on a planted bias. |
M1 |
| GATE-REFLEX-04 | Design comparison | Rank agreement of the baseline rung with trained specialists on a calibration set of 30 to 40 designs per morphology class and task family, thresholds fixed before data. A Kendall τ whose lower 95% bound is at least 0.7 lets the rung select; at least 0.5 lets it discard the bottom half; below 0.5 the rung removes only designs that fail outright and its results are labeled baseline-only: unranked (OPEN-REFLEX-2). |
M3 |
| GATE-REFLEX-05 | Known answer | PO-REFLEX-02; every command has a layer record; a planted infeasible program yields a certificate that checks exactly and a fallback event; a run that needs commands beyond the envelope is marked actuator-infeasible. |
M3 |
| GATE-REFLEX-06 | Design comparison | Sampling controller on the pilot design set: success at least 60% within three episodes; median at most 1 GPU-minute per 20 s quadruped episode and 6 per humanoid episode (the thresholds of BET-27); Kendall τ against specialists at least 0.5; FROB-REF-817 fires on a planted acceptance request that has no frozen planner record; the controlled world's digest does not change with the fork count. | M4 |
| GATE-REFLEX-07 | Known answer | Planted asymmetries (side slope, off-centre payload, front-only sensor, one-sided patch) remove exactly the expected generators; the two-latch fixture passes in law and fails tied; the event identity holds within tolerance on regular landings and is flagged on grazing ones. | M4 |
| GATE-REFLEX-08 | Known answer | Skill exact layer: both hand fixtures reproduce; 4 of 4 infeasible bodies are refused with certificates that check exactly; feasible witnesses replay; the {stable, fall} refinement is refused. | M4 |
| GATE-REFLEX-09 | Design comparison | Learned residual transport M: at least 18 of 20 feasible unseen bodies within 3 percentage points of tuned baselines with no adaptation, both cost denominators reported. | M8 |
| GATE-GAIT-01 | Known answer | Rimless-wheel fixed point and multiplier within 10⁻⁶; compass-gait multipliers within 10⁻⁶ of an independent high-precision computation; the planted transverse instability is found and the restricted computation is marked incomplete; PO-GAIT-01. | M4 |
| GATE-GAIT-02 | Known answer | On a 100-case suite: in at least 72 of 80 regular cases the gradient's cosine against central differences is at least 0.95; all 20 grazing and fold controls are labeled or refused; no case is refused for simultaneity alone; the bar fixture returns rest, v/4 and the sign of the coupling term. | M4 |
| GATE-GAIT-03 | Design comparison | Continuation succeeds on at least 50% of 10% design perturbations; median solve at most 10 minutes of CPU; Kendall τ of at least 0.5 between orbit cost of transport and specialists on the calibration set; the measured solve cost is recorded beside the published range. | M4 |
| GATE-SHIELD-01 | Known answer | Energy tier: the worked card reproduces; no exceedance over 10⁴ randomized contact sequences with μ in [0, 5], restitution in [0, 1] and every contact law; a budget with a planted unlisted spring is refused; a payload-only budget shows an exceedance. | M4 |
| GATE-SHIELD-02 | Known answer | PO-SHIELD-02 with zero box violations; the worked example gives (1, 0.5); a planted non-convergent multi-row instance invokes the backup; PO-SHIELD-03. | M4 |
| GATE-SHIELD-03 | Design comparison | 5,000 in-envelope and 5,000 shifted-world episodes with independent breach labels: in-envelope violations within the stated bound; shifted results reported separately; at least 85% of unshielded return; at most 5% added step cost. | M4 |
| GATE-SHIELD-04 | Known answer | Measured margin: empirical coverage at least 1 − α − 0.01 on 10⁴ held-out episodes of the calibrated family; a planted change of policy trips FROB-REF-457; a planted drift trips the alarm. | M4 |
| GATE-SHIELD-05 | Known answer | Reachable-set table M: zero false admissions in 5,000 in-envelope episodes on the braking fixture; uncovered outside the table. |
M8 |
| ✖ Kill gate | Trigger | Fallback, named in advance |
|---|---|---|
| ✖ KILL-BASELINE | GATE-REFLEX-02 fails (BET-06) | Baseline synthesis is limited to the morphology classes that pass; others report not-synthesizable and route to training |
| ✖ KILL-RESIDUAL | The pilot's falsifier holds, or at M4 the residual rung needs more than 50% of a specialist's steps on the corpus | The residual rung is removed; the baseline remains a behaviour-cloning warm start (§23.10) |
| ✖ KILL-MPC | GATE-REFLEX-06 fails after one repair cycle | The sampling controller is kept as a single-world CPU debugging aid; demonstrations come from the baseline and from specialists |
| ✖ KILL-GAIT | GATE-GAIT-03 fails | Orbits serve stability diagnostics and warm starts only; the atlas keeps supplying clocks and reward references; no design is ranked by an orbit |
| ✖ KILL-SKILL | GATE-REFLEX-09 fails | Keep the exact layer and its witnesses; drop the learned residual |
| ✖ KILL-ENERGY | An exceedance under the premises survives one repair cycle at GATE-SHIELD-01 | The claim is withdrawn; the monitor remains as an alarm of class Estimated |
| ✖ KILL-SHIELD | GATE-SHIELD-03 fails after one repair cycle | The energy tier stays; barrier rows become reward penalties and alarms; nothing is called a shield |
| ✖ KILL-TUBE | GATE-SHIELD-05 fails, or the pilot's falsifier holds | No tables are built; reduced-order barriers cover those subsystems |
22.19 Open questions
- OPEN-REFLEX-1. Whether the predictive controller runs on device. Owner: Reflex lead. Decided at M3. Default: the single-step program runs on every tier; the predictive controller runs on CPU tiers only until measured.
- OPEN-REFLEX-2. The rank-agreement thresholds of GATE-REFLEX-04. The candidate pairs for "select" and "discard" are (0.7, 0.5) and (0.6, 0.3); no published measurement decides between them. Owner: Reflex lead. Blocks GATE-REFLEX-04. Default: the stricter pair, and until the gate has run the baseline rung discards only.
- OPEN-REFLEX-3. The residual's authority ρ per channel. Owner: Reflex lead. Blocks GATE-REFLEX-05. Default: 0.3 of range.
- OPEN-REFLEX-4. Whether screening by the sampling controller needs a 4 ms planning step. Owner: Reflex
lead. Blocks GATE-REFLEX-06. Default: 20 ms, labeled in the
ControllerId, with results flaggedmodel-privileged. - OPEN-GAIT-1. When the referee route for orbits is built. Owner: Kinesis lead with upstream. Blocks nothing on the critical path. Default: compliant tiers only.
- OPEN-GAIT-2. The norm and tolerance for order sensitivity. Owner: Reflex lead. Blocks GATE-GAIT-02. Default: the mass norm and 1% of the pre-impact speed.
- OPEN-SHIELD-1. Where absorber envelopes come from before any bench data exist. Owner: Sinew lead with
Ground. Blocks any energy claim above
Estimated. Default:assumedcards andEstimatedclaims. - OPEN-SHIELD-2. The calibration level α and the per-episode violation budget. Owner: the project owner. Blocks GATE-SHIELD-03 and -04. Default: α = 10⁻³ and a budget of 10⁻³ per episode.
22.20 A lifted linear planning model for one soft-soil regime
What it is. M. A lifted linear predictor is a finite matrix acting on a dictionary of functions of the
state. As a global model for control through contact it is not built. A contact-rich hybrid flow has no
finite-dimensional invariant subspace of such functions, and size alone settles it: a quadratic dictionary on
60 states has 1 + 60 + 1,830 = 1,891 functions, so one dense map has 1,891² ≈ 3.58 × 10⁶ coefficients before
any mode jump is represented. One bounded experiment is kept, off the critical path. It covers a single regime
with no contact-mode change, a tracked or wheeled base driving on soft soil (§18.10), where the dynamics are
smooth and a rollout of the compiled model is comparatively expensive. There a lifted model is tried as the
planning model of the sampling controller of §22.8 and judged at equal tracking accuracy. Owner:
frob-control; bands come through the airlock from fs-surrogate. Verbs: control.lift (fit, calibrate,
report) and control.mpc --planner lifted:<id>.
FACT: fs-surrogate provides exact DMD, EDMD over a polynomial dictionary with ridge regularization, and
split-conformal forecast bands per horizon with an optional simultaneous adjustment; it states that the band is
its only accuracy statement and holds only inside the calibration distribution
(frankensim/crates/fs-surrogate/src/koopman.rs). Its EDMD is autonomous and has no input term. The
input-affine fit below is therefore ours, and adding it to fs-surrogate is an upstream change (Appendix F).
Model (DESIGN). The reduced state x has 7 entries: forward speed, yaw rate, the slip of each track, the sinkage of each track, and pitch. The inputs are u = (u_L, u_R), the two track torque commands. The dictionary ψ(x) holds every monomial of total degree at most 2: C(9, 2) = 36 functions. The predictor is bilinear in the lifted state and the input,
z⁺ = A·z + u_L·B_L·z + u_R·B_R·z, z = ψ(x),
fitted by ridge least squares on triples (x_k, u_k, x_{k+1}) from batch-cpu rollouts in f64. The state is
read back from the linear monomials and lifted again at every step. The three 36 × 36 maps hold 3,888
coefficients, and one planning step is three matrix–vector products, 3,888 multiply–adds.
Admission. The model carries its regime as data: the soil card with its declared ranges of pressure, slip
and sinkage (§18.10), a box of speed and yaw rate, a slope range, and the condition that every track sample
stays in contact. A simultaneous 95% forecast band over the planning horizon is calibrated on trajectories held
out from the fit. A query outside the regime, or a planner tick whose measured state leaves the box, is refused
and that tick is served by mpc-sample on the compiled model (refusal: lifted planner outside its calibrated
regime). Results carry screening: sampling-control and surrogate, are class Estimated, and inherit every
rule of §22.8, including that the controller never stands in for a trained policy as the evaluator of record.
Known answer and counterexample.
- Linear recovery. Data from x⁺ = 0.9·x with the dictionary {1, x} must return the map diag(1, 0.9) and the eigenvalue 0.9 to 10⁻⁹.
- A mode switch defeats a polynomial lift even when the fit is exact on its data. Take the reset map x⁺ = |x| and the dictionary {1, x, x²}. On the samples x ∈ {−1, 0, 1} the targets are {1, 0, 1}, which x² reproduces exactly, so the fitted predictor is x⁺ = x² with zero training residual. At x = 0.5 it predicts 0.25 where the truth is 0.5. A zero residual on the fitting set is no evidence. Only the held-out band is, and only inside one mode.
Ceiling, stated first. The lifted model can remove at most the planner's rollouts, N_s·H·N_it model steps per tick: 4 × 10⁴ to 2 × 10⁵ in the configurations of §22.8, against one plant step. Inside the controller the ceiling is that ratio. On a design cycle it is the share of wall time that the screening rung of §22.8 spends on soft-soil tasks in the phase ledger of §39, which is unmeasured and expected to be small. Nothing in M0–M7 depends on this item.
Pilot before funding. On batch-cpu only: fit and calibrate on the tracked skid-steer of §16.6 over one
soil card, and report band width against horizon. Falsifier: the simultaneous band at a 0.9 s horizon is wider
than the task's tracking tolerance (DESIGN default: 5% of commanded speed and 0.05 m of path offset). Equal
accuracy is then unreachable and the experiment stops.
Gate and kill gate.
| Gate | Claim | Threshold (TARGET) | Earliest |
|---|---|---|---|
| GATE-REFLEX-10 | Design comparison | Linear recovery to 10⁻⁹; the absolute-value control returns 0.25 against 0.5 and that fit is denied admission; empirical coverage of the simultaneous 95% band at least 0.93 on 1,000 held-out trajectories; closed-loop tracking error with the lifted planner within 10% of the compiled-model planner's on 100 episodes paired by world address; planner cost per tick at least ten times lower at that accuracy; every out-of-regime query refused. | M6 |
✖ KILL-SOILLIFT. Trigger: the pilot's falsifier holds; or more than 1% of calibration trajectories contain a
contact-mode change inside the declared regime; or the accuracy or cost clause of the gate fails after one
repair cycle. Fallback: mpc-sample on the compiled model, as §22.8 specifies; no lifted model is offered for
any other regime.
22.21 The pre-slip aliasing fixture and the sampled-data backup domain
This subsection extends §22.12.3 and §22.16. It defines the control named in BET-25, states when a monitored
assumption may count toward the bound of §22.12.3, and gives a two-mode sampled-data fixture whose backup domain is
checked by hand. Tier S: a known answer with negative controls. Owner: frob-shield, with the fixture registered in
frob-conformance. Verbs: shield.fit (the rule and its refusal) and shield.check (replay of the fixture).
The fixture. A cart on a level floor approaches a keep-out line at v₀ = 2.0 m/s with its wheels rolling and no
drag. A brake command asks for a deceleration of 5.0 m/s²; the floor delivers min(5.0, μ·g). There are two worlds:
A with μ = 0.8 (μ·g = 7.848 m/s², so the wheels hold and the cart decelerates at 5.0) and B with μ = 0.3
(μ·g = 2.943 m/s², so the wheels slide). The modes are stick and slip. The control period is Δt = 0.02 s. The
sensors are wheel speed, body acceleration and range s to the line.
| Quantity | World A | World B |
|---|---|---|
| Stopping distance v₀²/(2·decel) | 4/10 = 0.400 m | 4/5.886 = 0.6796 m |
| Observations before the first braking tick | Identical in both worlds, sample for sample |
Why the histories are identical. Before braking the traction demand is zero, so no contact law reads μ, and the delivered transcripts of the two worlds are equal. A monitor is a function of the delivered transcript. It therefore returns the same output in both worlds up to the first braking tick, whatever its design. The fixture asserts transcript equality as a precondition and is void on any tier where it fails.
Arm 1: a nominal controller with a slip monitor. The controller assumes μ ≥ 0.5 and brakes when s ≤ 0.45 m (the 0.40 m of world A plus 0.05 m). A monitor compares wheel speed with body speed and raises an alarm on slip; the backup is full braking. In world A the cart stops 0.05 m short. In world B the alarm is raised at the first braking tick and the backup is already running: the cart crosses the line at √(4 − 2·2.943·0.45) = 1.162 m/s and stops 0.6796 − 0.45 = 0.2296 m beyond it. For this stratum P(violation | alarm) = 1, and the bound of §22.12.3 is at least P(N_alarm ≥ 1) = 1, which says nothing. The alarm is correct and too late.
The rule this fixture enforces. For an assumption with a monitor and a backup, let t_d be the first tick at which
the delivered transcript of a violating world differs from that of every admitted non-violating world, and let ℓ be
the monitor's latency in ticks. The pair is timely if the state at t_d + ℓ lies in the backup's domain for every
violating world of the belief set. An assumption whose monitor is not timely may not be listed as monitored: its
whole range goes inside the quantifier of §22.12.1, so the rows hold for every value in the belief set. shield.fit
refuses a shield that lists such an assumption as monitored, and returns the first distinguishing tick, the state
there and the two changes that would admit it. Here they are: brake by s ≤ 0.7196 m, or limit the approach speed to
1.57 m/s, the largest v with v²/5.886 + 0.02·v ≤ 0.45.
Arm 2: the shield with the belief set inside the quantifier. Belief set μ ∈ [0.3, 0.8], so μ_min·g = 2.943 m/s². The backup is full braking, with domain
D = {(s, v) : s ≥ v²/(2·μ_min·g) + v·Δt}.
Invariance, by hand. If (s, v) ∈ D at a tick, coasting for one period leaves s − v·Δt ≥ v²/(2·μ_min·g). The filter
permits coasting only inside D and otherwise brakes. Braking therefore begins at a tick whose range is at least
v²/(2·μ_min·g), and the deceleration from then on is at least μ_min·g in either mode (5.0 in stick, μ·g ≥ μ_min·g
in slip), so the cart stops at or before the line for every μ in the belief set. At v₀ = 2.0 m/s the threshold is
0.6796 + 0.04 = 0.7196 m. The shield has no violation in either world. Its cost is visible: in world A the cart stops
between 0.28 m and 0.32 m short of the line, depending on where the sampling instant falls, where 0.05 m would have
sufficed.
Negative control: the continuous-time barrier in a sampled loop. Drop the term v·Δt and test h = s − v²/(2·μ_min·g) ≥ 0 at each tick. At a tick with s = 0.70 m, h = 0.0204 and the filter permits coasting. At the next tick s = 0.66 m, braking begins, and in world B the cart stops 0.6796 − 0.66 = 0.0196 m beyond the line. The overshoot of this mutant is at most v·Δt = 0.04 m. This is the inter-sample term L·Δt of §22.12.1 at its smallest size.
What the fixture does not show. It has one degree of freedom and an exact stopping law. It says nothing about the
tracking margin of §22.12.4 or about legged backups, whose domains are Tested only.
Gate. GATE-SHIELD-06 (Known answer, M4). On the referee and on the strict profile, over 1,000 paired episodes
with v₀ drawn from [1, 3] m/s and friction pairs from [0.3, 0.8]: the two transcripts are bit-equal before the first
braking tick in every pair. Arm 1 violates in the low-friction world in every episode whose stopping distance exceeds
its trigger range, with the overshoot within 1 mm of the closed form. Arm 2 has no violation in 10⁴ episodes. The
mutant without the inter-sample term produces at least one violation, none larger than v₀·Δt. shield.fit refuses
Arm 1's monitored-assumption listing and prints both admitting changes.
22.22 The sampling controller as the delivered controller
The delivered-controller lane F. This amends the rules of §22.8 and ADR-REFLEX-03. The rule "never the
evaluator of record" guards one substitution: scoring a design under a sampling controller and reporting the score
as if a trained policy had earned it. It does not cover the case where the sampling controller is itself what the
mission delivers. An accepted answer is a design, a controller and an evidence package (§2.1). When the controller
in that package is of kind mpc-sample, episodes run under it are the only evidence about it, and a refusal keyed
on the controller kind would leave the package with no evidence at all.
DESIGN: a task or mission may declare controller.delivered = mpc-sample. The declaration freezes a planner
record, whose digest is part of the ControllerId.
planner_record (frob.control.planner/1) DESIGN, fixes intent
kind mpc-sample
budget samples N_s · horizon steps H · iterations N_it model steps per tick = N_s·H·N_it
planning_step h_p · control rate
cost_terms content hash of the class-default or task-declared cost
planning_model ModelId rolled out by the planner · source of its parameters (nominal | estimated | plant)
state_source estimated (the filter of §22.3) | privileged
compute_target named computer · measured planning time per tick: median, 99th percentile · deadline
Episodes under a frozen planner record are admitted as the scoring source of acceptance, promotion and qualification for that record only, when four conditions hold. Otherwise the refusal of §22.8 stands.
- Not model-privileged. The planning model's parameters do not come from the plant's randomization draw, and
state_source = estimated. An episode flaggedmodel-privilegedis refused as acceptance evidence. - A plant the planner does not own. The plant runs on the tier the battery names and is promoted to the referee by the rule of §16.8, as for any controller. The planner keeps its own model and its own planning step.
- Measured latency.
compute_targetcarries a measurement. The planner's output is delayed by the measured 99th-percentile planning time, rounded up to whole control periods, and a tick that misses its deadline holds the previous plan. A record with a modeled or zero delay is refused for acceptance. - The same bar. The battery, the requirement verdicts and the failure bound are those of any other controller, and the cost is charged to the phase ledger of §39 in real GPU-seconds and CPU-seconds.
The verdict names what was accepted: accepted with delivered controller mpc-sample@<record digest>. It says
nothing about a trained policy for the same design, and a later change of controller voids it under §28.13.
Why condition 1 is not a formality. A cart at 2 m/s must stop before a wall on a floor whose friction is drawn from {0.2, 0.8}; stopping distance is v²/(2·μ·g). A planner that reads the plant's draw brakes at 1.019 m on the slippery floor and 0.255 m on the grippy one and never fails. The same planner holding the nominal μ = 0.5 begins braking 0.408 m from the wall and overruns by 0.612 m (1.0194 − 0.4077) on every slippery world, half of the draws, unless its estimator detects the floor in time. The privileged run reports zero failures in 299 episodes for a controller that fails every second episode.
Break-even, stated before any test. FACT arithmetic on the figures of §22.8. A 1% failure bound at 95% confidence needs 299 failure-free episodes; training once costs about 600 GPU-seconds for the quadruped and 3,600 for the humanoid. The lane is cheaper than training only below 600/299 = 2.0 GPU-seconds per 20 s episode for a quadruped and 3,600/299 = 12.0 for a humanoid. The published settings cost about 20 and 62 GPU-seconds per episode: 10 and 5 times over. At the throughput targets of §36 (2 M and 0.5 M world-steps/s) the break-even budgets are 4.0 × 10⁶ and 6.0 × 10⁶ model steps per episode, which at 50 Hz is 4,000 and 6,000 model steps per tick. One setting inside the quadruped budget is N_s = 64, H = 16 at a 20 ms planning step and N_it = 3 (3,072 steps per tick), against 65,536 for the published quadruped setting. HYPOTHESIS: a planner that small completes the reference tasks for some classes. FACT (Appendix H): controllers with about 30 samples run a real quadruped from a CPU; the published successes use hand-built swing-foot or stepping references; success in one study fell from 10 of 10 to 6 of 10 and 0 of 10 when the spline family changed; nothing was found for tracked robots or for a varied design corpus. Ceiling: no speed benefit at any published setting. The benefit is a capability: a design whose deliverable controller is a planner can be accepted at all, with its price on the ledger.
Owner, verbs and refusal. frob-control owns the planner record and the latency model; frob-arena admits the
episodes. control.mpc --freeze writes the record and eval.qualify consumes it. An acceptance request whose
scoring episodes are mpc-sample without a frozen, non-privileged, latency-measured planner record is refused,
under the existing refusal of §22.8, with the missing condition named.
| Gate | Claim | Threshold (TARGET) | First milestone |
|---|---|---|---|
| GATE-REFLEX-11 | Known answer; Fidelity | The cart: the privileged run is refused as evidence, and the nominal-model planner overruns by 0.612 m within 1% on the slippery draws. On the M3 corpus, every delivered-controller verdict on the fast tier survives promotion to the referee plant or is labelled tier-sensitive; the measured 99th-percentile planning time fits the control period on the named computer in at least 99% of ticks; the phase ledger prints the lane's cost beside the cost of training once |
M3 |
| Kill gate | Trigger | Fallback |
|---|---|---|
| ✖ KILL-MPCRECORD | More than 10% of delivered-controller verdicts reverse at the referee; or no named computer meets the deadline in 99% of ticks for any class on the pilot set | The rule of §22.8 as first written: screening and fallback only. A mission that requires a sampling controller returns provisional: controller-not-qualifiable with its screening episodes |
22.23 Orbits of a loop with no clock: the free period and the trivial multiplier
What it is. F The treatment of gaits whose closed loop is autonomous. §22.9.1 strobes a clock-forced loop over
one stride of a known period. A passive walker, a controller whose phase is reset by touchdown, and any loop that
reads no clock have no such period: the stride time is an unknown, and the equations of §22.9.5 and §22.9.6 are
singular on every such orbit. Owner: frob-gait. Verbs: gait.orbit, gait.continue, gait.gradient.
Why the clock-forced equations fail. On an orbit of an autonomous loop with period T*, a state displaced along the flow returns displaced by the same amount. The vector field f(x*) is therefore an eigenvector of the time-T* monodromy M with multiplier exactly 1, for every design and every controller. I − M is singular, the fold test of §22.9.5 fires on every orbit, and the solve of §22.9.6 has no solution or infinitely many. The multiplier carries no information about stability.
Mechanism. DESIGN.
-
Classification. A
ControllerIddeclares its phase source:clock,event-resetornone. A loop isclock-forcedonly underclock, andautonomousotherwise.gait.orbitrefuses a stroboscopic solve on an autonomous loop (refusal: fixed-period shooting on a loop with a free period). -
Section form. Where the stride has a repeatable transverse event with guard g(x) = 0, the orbit is a fixed point of the return map P_Σ on the section Σ = {g = 0} (C-16), of dimension one less than the state. Its Jacobian is the saltation product projected along the flow, D P_Σ = (I − f⁺·∇gᵀ/(∇gᵀ·f⁺))·M restricted to the tangent space of Σ. Its multipliers are those of M without the trivial one. Stability, folds and the gradient equation (I − D P_Σ)ᵀ·λ = ∇_Σ J are stated on Σ.
-
Bordered form. Where no event serves as a section, the unknowns are the state and the period, with a phase condition s(x) = 0, for example ⟨x − x_ref, f(x_ref)⟩ = 0. Newton solves
[ M − I f(x_T) ] [ δx ] [ φ_T(x) − x ] [ ∇s(x)ᵀ 0 ] [ δT ] = − [ s(x) ]The bordered matrix is nonsingular exactly when 1 is a simple multiplier and the phase condition is transverse to the flow. A fold is a singular bordered matrix. The same matrix, with the parameter derivatives on the right, gives dx*/dθ and the derivative of the stride time dT*/dθ; its transpose gives the adjoint for a metric.
-
What stays. The symmetry reduction of §22.9.2, the rule that stability is computed over the full reduced state, the derivative labels of §22.9.3 and the typed outcomes of §22.9.5 apply unchanged. The removed coordinates listed in a
GaitOrbitrecord gain the phase.
Known answers.
- A planar loop, by hand. ṙ = r·(μ − r), φ̇ = 1 in polar coordinates. The orbit is r = μ with period 2π. The time-2π monodromy has multipliers 1 and e^(−2πμ), so det(I − M) = 0. On the section φ = 0 the return map has the single multiplier e^(−2πμ), and the implicit-function solve gives dr*/dμ = 1. The bordered solve gives the same, with dT*/dμ = 0.
- The rimless wheel of §22.16. The state (θ, θ̇) has monodromy multipliers 1 and cos²(2α); on the post-impact section the multiplier is cos²(2α), which is 1/2 for eight spokes.
Pilot, ceiling, kill. Pilot: the planar loop, the rimless wheel and the compass gait through both forms. Falsifier: the two forms disagree on a nontrivial multiplier by more than 10⁻⁶. Ceiling: none on speed; the benefit is that orbit analysis applies to event-driven controllers at all, and that the stride time becomes a differentiable metric.
| GATE-GAIT-04 | Known answer. On the planar loop the section and bordered forms agree with the closed forms to 10⁻⁹ on the multiplier, on dr*/dμ and on dT*/dμ; a stroboscopic solve of the same loop is refused, and is not reported as a fold. The rimless wheel and the compass gait of GATE-GAIT-01 pass through the autonomous path with the trivial multiplier removed and listed. M4. |
| ✖ KILL-FREEPERIOD | Trigger: on the gait battery the section and bordered forms disagree beyond tolerance on more than 5% of autonomous orbits after one repair cycle, or the bordered iteration fails to converge on more than a quarter. Fallback: autonomous loops are analysed on an event section only; a loop with no transverse event is reported unsupported for orbit analysis and goes to finite-disturbance recovery tests, class Tested. |
22.24 Stability over admitted sequences of strides
What it is. F The stability object for a gait that can alternate between stride modes: a step up and a step
down, a left and a right asymmetric stride, a slipping and a gripping stride. Each mode has its own return-map
Jacobian. A multiplier test on each mode alone does not decide the alternation. Owner: frob-gait, with the shared
cone solver of §22.5. Verb: gait.orbit with a mode set.
Counterexample, by hand. On a two-dimensional section let A = [[1/2, 1], [0, 1/2]] and B = Aᵀ. Repeating either stride alone has both multipliers at 1/2. The two-stride product B·A = [[1/4, 1/2], [1/2, 5/4]] has an eigenvalue 3/4
- 1/√2, about 1.457. The alternating pattern amplifies a disturbance by (1 + √2)/2, about 1.207, per stride while both isolated tests pass.
Mechanism. DESIGN.
- The admitted language. The task's terrain grammar and the controller's mode guards define which mode may follow which: a directed graph on modes, registered with the orbit. A gait with one mode has the trivial graph and nothing here applies.
- Lower bound. Enumerate the admitted cycles up to a length L (DESIGN default 4, with a count cap) and take the largest ρ(product)^(1/length). A value above one is an instability with a witness word.
- Upper bound. Search for one matrix P ≻ 0 and a rate γ with A_σᵀ·P·A_σ ⪯ γ²·P for every mode σ, a semidefinite program of the size of the section, by bisection on γ. This is the construction §22.9.4 already uses for the orderings of a simultaneous impact. A common quadratic norm is sufficient and not necessary.
- Verdict.
stable-under-switchingonly if the upper bound is below one;unstable-sequencewith its word if the lower bound is above one;undecidedotherwise, with both numbers. The verdict is local, like every multiplier: it says nothing about the recovery basin, which is tested with finite disturbances.
For the counterexample the lower bound at length two is 1.207, no common contraction exists, and the verdict is
unstable-sequence with the word (A, B).
Pilot, ceiling, kill. Pilot: the counterexample, then a quadruped trot over alternating risers on the compliant
tier, with the bracket compared against 200 sampled admitted sequences of 50 strides. Falsifier: a sampled sequence
grows under a stable-under-switching verdict, or the bracket is undecided on more than half of the multi-mode gaits
of the battery. Ceiling: none on speed; it removes one class of false stability labels.
| GATE-GAIT-05 | Known answer. The counterexample returns the eigenvalue 3/4 + 1/√2 to 10⁻¹², the per-stride bound 1.2071 and the word (A, B); a pair of commuting contractions returns stable-under-switching with a common matrix that the checker verifies in exact rationals; a mutant that tests each mode alone labels the counterexample stable and fails. M4. |
| ✖ KILL-STRIDESEQ | Trigger: the pilot's falsifier holds after one repair cycle, or cycle enumeration exceeds its budget on the reference gaits. Fallback: no switching verdict is issued; multi-mode gaits are tested by finite-disturbance recovery over sampled admitted sequences, class Tested, and the per-mode multipliers carry the flag single-mode-only. |
22.25 Sampled damping, and the work of a held command
What it is. S The sampled-data condition under which the damping-injection backup of §22.12.3 is passive, and
the rule for accounting the work of any command held over a control period. In continuous time u = −K_d·y never adds
energy. A command computed at a tick and held for a period h can. Owner: frob-shield (fitting), frob-control (the
backups), frob-chronicle (the ledger). Verbs: shield.fit, shield.check.
Counterexample, by hand. A free 1 kg slider moves at +1 m/s. A sampled damper with gain 300 N·s/m reads the speed and holds −300 N for 10 ms. The speed ends at −2 m/s. The kinetic energy rises from 0.5 J to 2 J: the damper has done +1.5 J of work. The power at the start of the interval times its duration predicts −3 J.
The condition. DESIGN. For the port model of C-12 with reflected inertia M̄ at the collocated outputs and no other force during the hold, one tick maps the output rate y by I − h·M̄⁻¹·K_d, and the stored energy changes by −(h/2)·yᵀ·(2·K_d − h·K_d·M̄⁻¹·K_d)·y. Therefore:
- stored energy does not increase at the ticks exactly when 2·M̄ − h·K_d ⪰ 0;
- the rate does not reverse in any direction exactly when M̄ − h·K_d ⪰ 0;
- with one tick of actuation latency the scalar loop v′ = v − κ·v_prev, κ = K_d·h/M̄, is stable only for κ < 1, half the bound without latency.
shield.fit computes the bound from the least reflected inertia over the design's mass ranges and the control period
of the task, checks it in exact rationals, and fits the backup gain inside the no-reversal bound by default. With
latency or with other stored energy in the loop it uses the exact one-tick map of analysis A-22 and requires a
contraction in the energy norm. A gain outside the bound is refused (refusal: sampled damping gain above the passivity
bound). A brake or a shorted winding is a physical damper and carries no such bound; the backup's card says which it
is.
The work of a held command. DESIGN. For a force F held over h on an inertia m moving at v, the work is F·v·h + F²·h²/(2m). The second term is never negative, so the power at the tick times the period always understates the work done on the plant. Any budget that debits predicted work ahead of the motion (for example an allowance on the work of the learned residual of §22.7, or a monitor that predicts the drive term of §22.11) reserves this bound with the inertia at its least admitted value, or locates the instant at which the allowance is exhausted and changes the command there. The ledger of §29.9 is unaffected: it integrates port power with the velocities of each physics step.
Known answers.
| Case | Value | Mutation that must fail |
|---|---|---|
| The slider, κ = 3 | −2 m/s; +1.5 J | Work taken as initial power times duration: −3 J |
| κ = 2 | The speed reverses to −1 m/s; energy unchanged | — |
| κ = 1 | Stops in one tick; −0.5 J | — |
| κ = 3/2 with one tick of latency | Diverges: the map has spectral radius √(3/2) | A bound of κ < 2 applied to the delayed loop |
| GATE-SHIELD-07 | Known answer. The four rows are exact. shield.fit refuses 300 N·s/m at 10 ms on an inertia of 1 kg and admits 100 N·s/m. Over 10⁴ random triples (M̄, K_d, h) with 2·M̄ − h·K_d ⪰ 0 the ledger's stored energy at ticks never increases, and each of 10³ planted violations shows an increase. The continuous-time identity remains a theorem card. M4. |