What exists, what is missing, what this is
5,502 words8 subsections4 tablesPlan version 1.1
On this page8
This chapter is an audit before it is a thesis. It states what the strongest existing systems already do, what they
document as absent, what this plan therefore does not claim, and only then what FrankenRobotics is and where it may
differ. The incumbents were read in their own files at pinned versions: MuJoCo and MuJoCo Warp 4.0.1, Newton 1.8.0,
Warp 1.19.0, Isaac Lab 3.0.0 with Isaac Sim 6.2.0, Genesis 1.4.7, Drake 1.58.0 and ManiSkill 3.0.1 (FACT, Appendix H).
No incumbent GPU simulator was run (the one executed incumbent measurement is CPU MuJoCo on this Mac, §36.3 row 13):
a capability listed here is documented, not demonstrated, and where a search found nothing the text says "not found",
which is absence of evidence and not evidence of absence. The audit is repeated with execution at M0, when the
incumbents are installed on trj and their strongest warm configurations are frozen (§39).
1.1 What the strongest existing systems already do
Every row is a FACT about its source (Appendix H). The third column is the consequence this plan draws.
| Capability | What is documented or published | Consequence for this plan |
|---|---|---|
| Training speed | A quadruped joystick policy trains in about 410 s on one RTX 4090 and about 325 s on two, compile time included (MuJoCo Playground, a figure read by eye, ± 20 s; the text says "within 5 minutes" on two). Humanoid policies train in under 15 to 30 minutes on two RTX 4090. A default rough-terrain humanoid run takes about 1 h on one RTX 4090 (derived from Isaac Lab's published configuration and frame rate). | The training rung is a parity requirement. The 15-minute target of version 0.2 was slower than the published run on the same card and is replaced (§36). Narrative denominators are minutes to about an hour. |
| Physics throughput | About 2.55 M humanoid (G1) world-steps/s at 8,192 worlds on an RTX 6000 Ada on flat ground, and 1.68 M on a heightfield (MuJoCo Warp nightly benchmark). Physics-only 4.8 M physics steps/s for a quadruped at 4,096 environments on an RTX 5090, with 71% of the full environment step spent outside physics (mjlab nightly). CPU MuJoCo runs the classic humanoid at 650 k steps/s on an M3 Max and 1.8 M on 64 cores. | Throughput targets are restated as ratios to the incumbent measured on trj under the disclosure block (§36). No order-of-magnitude headroom per step is assumed anywhere. |
| Determinism | Opt-in bitwise run-to-run determinism on the same GPU model, driver, toolkit, build, model, inputs and step sequence: Newton 1.8.0 for five solvers, and Isaac Lab 3.0.0 on PhysX for rigid bodies and articulations. | Deterministic GPU replay as such is not a differentiator. §1.2 says what is documented as absent. |
| Different robots in one batch | ManiSkill 3.0.1 merges articulations with different link and joint structures, padded to the largest. Newton 1.8.0 documents heterogeneous worlds for two solvers and labels the support experimental. One paper trains a single policy across three-, four- and five-legged bodies and reports the mixed batch at 86 k samples/s at 2,048 environments, 0.86 of the harmonic mean of its homogeneous rates. | Not a capability differentiator. Only two measured numbers can differentiate (§17, N11). |
| Per-world parameters and assets | MuJoCo Warp 4.0.1 documents batched model fields and per-world assets. | Per-world parameter variants within one topology are parity, and are S here. |
| Structural editing | CPU MuJoCo's mj_recompile "updates an existing mjModel and mjData pair in-place, while preserving the simulation state", including adding or removing bodies during simulation. On the GPU a model edit needs a host compile and a new graph capture. |
Editing structure is prior art on the CPU. No edit-to-first-GPU-step latency is published, so that is a number to measure, not to assume. |
| Forking state | CPU MuJoCo copies integration state so that source and copy step identically, and its batch API is bit-identical at any thread count. No GPU fork verb was found. | Counterfactual replay is established on the CPU. |
| Control with no training | Sampling-based predictive control with about 30 sampled trajectories and three iterations on a desktop CPU runs a real quadruped at 50 Hz, and a simulated humanoid with 60 to 70 samples. A GPU variant uses 2,048 samples. Each published controller uses cost weights chosen per task and robot. | "No training needed" is not a claim of this plan. The roles of baseline synthesis and sampling control are separated (BET-06, BET-27). |
| Differentiability | Available in MJX on JAX, Brax, Newton and Genesis; gradient quality through contact was not audited. | Not a differentiator. Contact gradients remain deferred (§6.2). |
| Actuator physics | MuJoCo's DC-motor actuator has optional inductance, cogging torque, thermal resistance variation, LuGre friction and a winding-temperature state, on the CPU and, in MuJoCo Warp 4.0.1, on the GPU. | Electrical and thermal actuator modelling is parity, not novelty (GOAL-12). |
| One kernel source, several backends | Warp compiles one kernel source for CPU and CUDA. Genesis's compiler lowers one source to CUDA, ROCm, Metal, Vulkan, x86 and ARM64. | Conceded. N1 claims only the derivation and the stored disagreement record. |
| Several solvers on one model | Newton documents a solver matrix over one model; Isaac Lab 3.0 introduces a multi-backend architecture over PhysX and Newton. | Conceded. N3 claims only the promotion rule. |
| Sequential comparison | Anytime-valid sequential comparison of two robot policies on binary, partial-credit and bounded continuous outcomes is published (arXiv 2603.13616). An earlier sequential test needs a maximum sample size fixed in advance (arXiv 2503.10966). | N4 claims only the application to paired design comparisons inside the studio. |
| Certificates | Drake 1.58.0 certifies collision-free polytopes in configuration space with sum-of-squares programs. | Kinematic certificates are prior art. N5 is narrower and says so. |
| Apple GPU | Genesis 1.4.7 lowers kernels to Metal. No published rigid-body throughput on Apple silicon was found. | The Metal tier is a number, not a first (BET-09, N10). |
| Interior scenes | Homes and kitchens (one project ships 2,500 or more kitchen scenes; others generate interiors from text or procedurally, with articulated fixtures). Two published retail-store generators. A warehouse builder that turns a 2D plan into a scene and has a headless interface. | GOAL-06 claims the grammar system and the two archetypes for which no generator was found (N12). |
| Agent access | Isaac Sim 6.2.0 ships an MCP server for documentation lookup. A community server exposes 42 tools that drive the simulator. | MCP access to a simulator is prior art. What remains is in N9. |
| One policy across bodies | One policy trained on 16 legged robots (1.6 × 10⁹ steps), on 50 robots (5 × 10⁹ steps, about 40 A100-hours) and on about 1,000 embodiments (over 2 × 10¹² steps on 160 GPUs for five days), with zero-shot transfer to withheld robots that resemble the training set. | Family policies per morphology class (BET-15). The third compute scale is out of reach of two cards. |
| Population co-design | Sharing specialist policies across candidates cut co-design wall-clock by 67.4%, about 3.1×, in design spaces of at most four dimensions. Clustering similar morphologies explored about 780 times more designs with 78% fewer simulation steps, at slightly lower best fitness. | Order-of-magnitude gains are published for designs evaluated per unit of simulation, not for the time or the quality of the best design. |
| Agent-driven design loops | Published loops spend about an hour of reinforcement learning per round of 16 candidates on eight GPUs, and train 80 to 1,600 candidates per campaign at 7.5 to 30 GPU-minutes each. | This is the denominator of the ten-fold test (§2). |
| Co-design theory and search | Grammar-based robot search, monotone co-design theory and language-model design loops all exist. | None is open ground. |
1.2 The fossils, corrected
Version 0.2 listed eight "fossils" of robot simulation. The incumbent audit showed that several described incumbents wrongly. The table below keeps only what an incumbent documents as absent or limited, quotes or cites the documentation, and names what this plan does about it and what must pass before the plan may say so.
| Stage | What incumbents document as absent or limited | Where | What this plan does, and the condition on claiming it |
|---|---|---|---|
| Replay | Bitwise agreement between a CPU run and a GPU run. Results "usually differ ... between the CPU and CUDA backends"; the kernel compiler defines only run-to-run and GPU-to-GPU determinism modes. | Newton 1.8.0 determinism guide; Warp 1.19.0 | Any world is re-run alone on the CPU from its address, bit for bit where the device reaches class D1 (N6). A hypothesis until GATE-SPARK-01 and GATE-SPARK-02 pass. |
| Replay | Independence of a world's trajectory from the batch it ran in. One solver's own tests hold a world equal to its isolated run only to 10⁻⁵–10⁻⁴; no guarantee is documented. | Newton 1.8.0 test suite | A batch-composition invariance test is part of the determinism program (§35). Until it passes, replay of a world taken out of its batch is a tolerance replay. |
| Determinism | GPU determinism in one leading engine: "Is MJWarp on GPU deterministic? No." | MuJoCo Warp 4.0.1 documentation | A determinism class is typed on every result and tested (GOAL-20). |
| Design | A topology change inside a running GPU batch. One engine keeps one model per batch and varies only fields and assets per world; one binding requires identical links and joints; heterogeneous worlds elsewhere are experimental, and "not all solvers are able to simulate them". | MuJoCo Warp 4.0.1; Isaac Lab 3.0.0; Newton 1.8.0 | Per-world model variants inside padded capacities and size buckets (§17). Claimed only as two measured numbers (N11). |
| Energy | Constraint energy: "potential and kinetic energy in constraints is not accounted for". | MuJoCo 4.0.1 XML reference | An energy ledger that closes to a stated residual with dissipation attributed per contact (N8), claimed only once the residual is measured. |
| Apple GPU | Metal acceleration in the Warp and JAX stacks: macOS wheels "support CPU execution but not Metal acceleration"; "macOS (CPU only)"; "not supported on Mac/OSX GPU". | Warp 1.19.0; Newton 1.8.0; JAX installation notes | A Metal tier measured against the one incumbent that does lower to Metal, on the same machine (N10). |
| Symmetry | Automatic symmetry: the reference library takes each robot's symmetry group from a per-robot configuration file. | MorphoSymm documentation | The symmetry group is computed from the design; a declared symmetry is an assertion that is checked (§13, §28). |
| Control | Automatic cost synthesis: each published sampling controller uses cost weights chosen per task and robot. | The controller papers (Appendix H) | A baseline controller synthesized from the model per morphology class (N7), conditional on BET-06. |
| Statistics | Paired comparison of designs on common random numbers. The published sequential tests compare policies, and one of them needs a fixed maximum sample size and binary outcomes. | arXiv 2503.10966; arXiv 2603.13616 | The published tests, applied to paired design comparison on common world addresses (N4). |
| Agents | Execution under budgets. The vendor MCP server is lookup, not execution, and neither it nor the community server documents compute budgets, provenance or a typed refusal taxonomy. | Isaac Sim 6.2.0 documentation; the community server's description | Every verb carries a compute budget, provenance and a typed refusal (N9). |
| Transfer | Policies for bodies far from the training set: the authors of the multi-embodiment results report that transfer "completely out of the training distribution is still problematic". | The multi-embodiment papers (Appendix H) | Family policies per class with an out-of-distribution check and a fallback (§23). |
| Gradients | Differentiation on the newer GPU back end of one engine: "not currently available in MJX-Warp". | MuJoCo 4.0.1 MJX documentation | Nothing. Listed for completeness; contact gradients are deferred here too. |
A second, weaker list holds what was not found after a bounded search. Each item is an absence of evidence. None is called a gap in an incumbent, and each has a line in the M0 audit that tries again with the software installed.
- Checked certificates attached to the kernels of a GPU engine, or enclosures of simulated quantities.
- Pre-simulation design analyses built into an engine (static sizing toolkits for serial arms exist outside engines).
- A promotion rule between solvers or tiers, or a cross-tier disagreement ledger.
- A generator for manufacturing plants or for commercial kitchens.
- An integrated GPU tool for rare-event or adversarial failure search (research methods exist).
- Structural analysis, topology optimization or fatigue inside an agent's design loop (one co-design framework has an automated finite-element filter).
- A published latency from a topology edit to the first GPU step.
- A published end-to-end phase ledger of a design loop on the late-2026 software stack.
- A published rank agreement between a shared policy and per-design specialists.
- A published ten-fold gain in end-to-end time to a qualified legged-robot design against a comparable baseline.
- A rigid-body throughput figure for the one Metal-capable incumbent on Apple silicon.
1.3 What this plan does not claim
The following statements appeared in version 0.2 or follow from a careless reading of it. They are withdrawn (DESIGN; the reason for each is the row of §1.1 or §1.2 that contradicts it).
- Deterministic GPU replay as such. Two incumbents document opt-in bitwise run-to-run determinism.
- "The CPU version is a different code base." One kernel compiler emits CPU and CUDA code from one source, and one engine runs on its CPU device.
- Different robot topologies in one GPU batch, as a capability. One environment suite and two solvers of one engine document it.
- The combination of mixed topologies, determinism and articulated contact as an incumbent gap. One experimental solver documents all three, and its tests exercise the combination on small chains.
- Electrical and thermal actuator modelling as a novelty.
- "No training needed" for baseline control.
- Anytime-valid comparison as a novelty of the field.
- One kernel source for several backends, and simulation on several tiers, as novelties.
- Apple-GPU support as a first.
- Interior scenes for homes, kitchens, retail stores and warehouses as a novelty.
- An agent interface to a simulator as such.
- The humanoid throughput target of version 0.2 as parity, and its note that no absolute incumbent figure was found. Published figures exist and the target was about five times below one of them.
- Any ten-fold claim that rests on per-step throughput, on sampling control as the evaluator, or on a published ten-fold in time to a qualified design. None is supported.
- "Morphology is fixed before the simulator starts." State-preserving recompilation on the CPU contradicts it.
- That an imported model with an impossible inertia "loads without complaint" everywhere. One incumbent's compiler rejects physically impossible inertias.
- That an incumbent loop teaches an agent its mistake "six hours later". Published training takes about seven minutes for a flat-ground quadruped task, under 15 to 30 minutes for humanoids, and about an hour for a default rough-terrain humanoid run.
1.4 What is already solved, and what nobody has composed
Each of the following exists and works (sources in Appendix H):
- Linear-time articulated-body dynamics in spatial algebra (Featherstone), and its Lie-group formulation (Park, Bobrow, Ploen).
- Convex, compliant contact that scales to thousands of GPU worlds (MuJoCo's formulation in MuJoCo Warp and Newton; Drake's SAP and its successors on the CPU).
- Lie-group and variational integrators with discrete conservation laws (Iserles et al.; Marsden and West; Lee, Leok, McClamroch).
- Invariant filtering for legged state estimation (Barrau and Bonnabel; Hartley et al.).
- Morphological symmetry groups and equivariant policies (Ordoñez-Apraez et al.), and the symmetry classification of gaits (Golubitsky, Stewart, Buono, Collins).
- Monotone co-design with guaranteed fixed points (Censi), grammar-based robot search, and trajectory optimization with design variables.
- Anytime-valid sequential inference, already implemented in FrankenSim (
fs-race,fs-eproc), and published for robot policy comparison. - Sampling-based predictive control of legged robots without offline training.
- One policy across a class of bodies, and policy sharing across design candidates.
- Certified collision-free regions of configuration space; support-region and wrench-cone feasibility as linear programs (Bretl and Lall; Orsolino et al.).
- Exact family answers on small plants by standard means. In the pilots behind this version, a lifted robust linear program, a projection, an SMT solver and quantifier elimination each returned the same complete family answer as a specialized method on every fixture tried (FACT, Appendix K).
- Adaptive stress testing and importance sampling for rare failures; delta debugging; conformal prediction; topology optimization.
What was not found is their composition: one design identity and one evidence ledger carried across static analysis, simulation tiers, statistics, family queries, adversarial search, calibration and solid mechanics, in one typed algebra, so that a statement proved or measured at one layer is usable at the next and is withdrawn when an edit invalidates it. FrankenSim's thesis is that such a typed algebra collapses an archipelago of tools into a continuum (FACT: FrankenSim plan §1). FrankenRobotics applies that thesis to robots and takes it one step further: the algebra is not only the type system, it is the source code of the kernels.
That the composition is worth more than its parts is a HYPOTHESIS, and it is the central one. Partial precedents exist (a co-design framework with a structural filter, a static sizing toolkit for arms, anytime-valid policy comparison). Almost every mechanism in this plan can in principle be hosted by an incumbent engine with a layer on top (H-19, ACTIVE). The only test of the composition is the end-to-end comparison against a composed incumbent baseline in §39, and §4 (BET-14) says what the program does if that baseline ties.
1.5 What FrankenRobotics is
Six products that share one model, one evidence ledger and one verb registry.
- A proving ground (Arena, Cortex). A vectorized, device-resident environment API over batched physics. It runs legged, wheeled, tracked and hybrid robots across generated terrain and generated or ingested interiors: homes, apartments, offices, warehouses, manufacturing plants, retail stores, commercial kitchens. Many different robots share one batch (Phylum). Missions are stated as requirements, and a result is qualified in a lane that the discovery process cannot read.
- A design studio for agents (Atelier, Tether). A CLI and an MCP server with verbs to create and edit a robot design, analyze it without simulation, simulate it, compare it against another design with statistical honesty, read structured telemetry about why it behaved as it did, optimize it, and ask which earlier verdicts survive an edit.
- A mission-query engine (Augur). Given a mission, a catalogue of parts, a controller class and a description of what the robot can observe and when, it returns the set of designs that can do the job: an inner set with witnesses, an outer set with exclusions, the unresolved cells between them, and the cheapest experiment that would shrink the gap. The body, a sensor and the building are all admissible answers.
- An adversary and a counterfactual debugger (Nemesis, Autopsy). A search for the worlds that break a design, bounded by physical legality and by a witness that the design's own body could have coped; a reduction of each failure to a minimal counterexample world; valid bounds on rare failure rates; and an explanation of a failure obtained by intervening on forked replays, reported with the stability of the attribution.
- A path to reality and to the workshop (Ground, Foundry). Calibration of the model against imported measurements with predictions frozen in advance, reality bands issued only inside a calibrated family, and build packets (parts, geometry, tolerances, fatigue margins) that carry machine-readable flags saying they authorize nothing.
- A mathematical instrument (Marrow, Lemma, Spark). A Lie-algebraic formulation of design, simulation and control in which the mathematics is load-bearing: it generates kernels, supplies test oracles, prunes designs before simulation, and, where a checker can accept it, carries proofs and portable certificates that an offline auditor re-checks.
Five nouns recur and are defined in Appendix A. A design is a robot's intrinsic properties and connections. A world is geometry, material fields, lighting, moving objects and their behaviour. A task is initial states, observations, actions, goals, terminal events and metrics. A mission is a task family with requirements and the quantifiers that relate designs, controllers and worlds. A study binds versions of these to controllers, random streams, numerical settings, hardware profiles and one declared analysis rule.
1.6 The differentiators, worded as the audit permits
Each wording below is the most the audited evidence supports, and it is the only wording this plan, the product's documentation and any public description may use. A condition in brackets appears with the claim until the named gate has passed. The capabilities of products 3 to 5 in §1.5 are claimed as functions of the product and not as firsts; their only comparative claim is N13.
| ID | Differentiator | Permitted wording | Bets, goals |
|---|---|---|---|
| N1 | Derived kernels with a disagreement record | "Kernel arithmetic is derived once and emitted for a CPU reference, CUDA and Metal, and the measured cross-backend disagreement is stored per kernel. Compiling one kernel source for several backends is prior art (Warp, Genesis)." | BET-01, BET-02 |
| N2 | Static analyses as verbs | "The first shipped suite of pre-simulation design analyses callable by an agent. Nothing in an incumbent's engine prevents it from adding them." | BET-05 |
| N3 | Promotion discipline | "A fast-tier result stays provisional until it has been re-evaluated on a higher tier and the difference recorded. Engines with several solvers or back ends exist; we found none with a promotion rule." | BET-03, GOAL-16 |
| N4 | Paired design comparison | "Published anytime-valid tests, applied to paired design comparisons on common world addresses." Cites arXiv 2603.13616 and arXiv 2503.10966. | GOAL-15 |
| N5 | Kernel certificates | "Named kernels carry offline-checkable certificates of algebraic identities. We found no incumbent engine that does this. The certificates say nothing about floating-point results." [Value unproven until BET-11.] | BET-11, GOAL-17 |
| N6 | CPU replay of a GPU world | "Any world can be re-run alone on the CPU from its address. Where a device reaches determinism class D1, the CPU run matches the GPU run bit for bit. Incumbents document bitwise replay only for the same batch on the same GPU." [Hypothesis until GATE-SPARK-01 and GATE-SPARK-02 pass, and until the test that a world's trajectory does not depend on the batch it ran in passes.] | GOAL-10 |
| N7 | Synthesized baseline controller | "A controller is synthesized from the model in seconds for each morphology class, with no per-robot cost tuning." [Conditional on BET-06.] The words "no training needed" are not used. | BET-06, GOAL-13 |
| N8 | Energy ledger | "A per-actuator and per-contact energy balance that closes to a stated residual. Incumbents expose the inputs; MuJoCo documents that it does not account for constraint energy." [Only once the residual is measured.] | GOAL-11 |
| N9 | Agent surface | "Every verb carries a compute budget, provenance and a typed refusal." MCP access to a simulator is prior art. | BET-10, GOAL-18 |
| N10 | Metal tier | "Measured quadruped throughput on an M5 Ultra, reported against Genesis on Metal on the same machine and against one RTX 4090." A number, not a first. | BET-09 |
| N11 | Mixed-topology batches | Only as two measured numbers against Newton with Kamino or Steel on trj: distinct designs times steps per second, and time from an edit to the first GPU step. The published 86 k samples/s for a mixed batch of three-, four- and five-legged bodies is the baseline to beat. A third candidate is a property, not a number: run-to-run replay and forks inside a mixed batch. It is claimed only after the incumbent's deterministic solvers have been run on a heterogeneous batch and found unable to do it. |
BET-18, GOAL-22 |
| N12 | Scenes | "Seven interior archetypes from one grammar system, including manufacturing plants and commercial kitchens, for which we found no generator." | GOAL-06 |
| N13 | The composition | "One design identity and one evidence ledger across static analysis, simulation tiers, statistics and solid mechanics." A hypothesis whose only test is the end-to-end benchmark against a composed incumbent baseline. | BET-13, BET-14 |
1.7 Why the suite, and why now
FACT, from inspection of the sibling repositories on 2026-11-21 (FrankenSim at commit 483d09c682f2):
- FrankenSim already contains the hard referee-grade pieces. SE(3) and SO(3) with exponential, logarithm, adjoint,
coadjoint and Jacobians (
frankensim/crates/fs-ga/src/lie.rs); Featherstone forward and inverse dynamics for floating-base trees (fs-mbd/src/articulated.rs); a 29-degree-of-freedom Unitree G1 model with provenance (fs-mbd/src/robot_models.rs); a semismooth-Newton Signorini–Coulomb contact solve and certified continuous collision detection (fs-contact); a discrete Euler–Poincaré step on SE(3) with its adjoint (fs-time); CMA-ES, Bayesian optimization and MAP-Elites (fs-dfo,fs-bo,fs-archive); max-plus algebra (fs-tropical); anytime-valid racing and e-processes (fs-race,fs-eproc); units, intervals, evidence and a ledger. - It also contains pieces version 0.2 did not use. Multilevel Monte Carlo, control variates, Gaussian importance
sampling and Shapley effects (
fs-uq); value-of-information ranking (fs-voi); sum-of-squares programs and interval verification (fs-sos); dimensional-analysis regimes (fs-regime); an out-of-process package checker (fs-checker); Floquet analysis of a supplied monodromy (fs-spectral/src/floquet.rs); topology optimization and tolerance allocation (fs-topopt,fs-toleralloc); rods and fatigue (fs-solid, with the rainflow counter infs-material); port-Hamiltonian interconnection (fs-phs); film and tribology models (fs-tribo,fs-lbm). No splitting or subset-simulation code exists there. - FrankenSim does not contain what a gym needs. No batched environments, no GPU path in its workspace, no terrain
or interior scene representation for robots, no robot sensors, no MCP server. Its articulated contact step refuses
more than 21 contacts (
fs-contact/src/sustained/solve.rs:MAX_CONTACTS = 21) and it has no complete persistent contact time stepper (fs-contact/CONTRACT.md). Its accelerator doctrine says "FrankenSim does not ship an accelerator backend" and "The CPU reference is permanent" (frankensim/docs/ACCELERATOR_DOCTRINE.md). - The suite has proven the GPU pattern outside FrankenSim's workspace. One kernel source compiled for CUDA through
cudarcand NVRTC and for Metal through a prelude of macros, checked against a CPU emulation (frankensim/apps/instrument-lab/render-gpu-piano/src/{kernel.h,cuda.rs,metal.rs,emulate.rs}), and a generic Metal compute gateway whose documentation names "a physics integrator" as an intended consumer (frankentorch/crates/ft-kernel-metal/src/compute.rs). The gateway accepts at most 31 buffer arguments (MAX_ARGUMENTS = 31). - The MCP substrate exists.
fastmcp_rustis a cancel-aware MCP framework built on asupersync contexts and checkpoints, with tool, resource and prompt surfaces (fastmcp_rust/crates/fastmcp/src/lib.rs). Its own documentation says support for the 2026-07-28 protocol revision is under implementation and not yet claimed, sofrob-mcpis admitted through INT-FMCP-01, and KILL-MCP names its fallbacks, the last of which is the CLI alone (§32, §33). - The limits of the siblings are known and shape the design. FrankenTorch tensors are host memory and its device
tags are metadata, so the learner runs in this project's own device kernels (§23).
fs-qtyneeds the nightly featuregeneric_const_exprs(frankensim/crates/fs-qty/src/lib.rs), and most FrankenSim physics crates depend on it, which ties the toolchain to nightlies on which that feature compiles: a named risk tested at INT-TOOLCHAIN-01, with a small stablefrob-unitscrate as the documented escape hatch. The siblings are on four different pinned nightlies today (2026-07-06,2026-08-20,2026-08-25,2026-08-31). That is the starting state. The whole build graph uses one toolchain T, chosen at INT-TOOLCHAIN-01, and the siblings that are not on T are moved to it in their own repositories by a toolchain train (§33.13). FrankenLean's tactic crate (franken_lean/crates/fln-anvil, 1,423 lines) implements a discrimination tree and theomegaprocedure, and its other tactics are stubs; FrankenSymPy's real-root machinery is capped (8 variables, degree 256, 4,096 bits, zero-dimensional ideals), and it has no homotopy solver. Proof routes (§12) and synthesis lanes (§25) are sized to those caps.
FACT (to recheck), reported by review and not yet reproduced by this program at a pinned commit. Each becomes an admission test and an upstream work item (§33, Appendix F) before anything depends on the component:
- Three errors in
fs_g1_train::ppo_update(frankensim/crates/fs-g1-train/src/ppo.rs): for a negative advantage the clipped surrogate is applied on the wrong side of the ratio; the critic gradient into the shared bias is zero where the squared-error objective gives a non-zero value; and concatenating two independently reset one-step episodes moves a frozen policy's second ratio from 1 to 1.582459. - A wrong joint law in the action sampler of the same file (
fs_g1_train::gaussian_action): each marginal is correct, and adjacent channels agree in sign with probability 0.6655 where independence gives one half (§23.25, §33.17). - A near-straight integration defect in
fs-mbd's planar drive (frankensim/crates/fs-mbd/src/planar_drive.rs): at a heading of 32 rad with wheel speeds that differ by 6 × 10⁻¹³ m/s the routine returns zero translation where the true one-second displacement is about (0.834, 0.551) m. It must be qualified against an independent SE(2) exponential before it is used as an oracle. - A stride-convention trap between
franken_numpyandfrankentorch: an unconverted stride reads 12 where 3 is correct. - A fixed-witness assumption in the sustained articulated contact path of FrankenSim: it assumes fixed material witnesses and a fixed frame, and rolling around a sharp lip changes both. The referee's transactional step therefore migrates witnesses (§16).
These are five defects or limits (§33.11, §33.17). Beside them stands one convention hazard, confirmed by reading:
fs-mbd::Wrench holds a world-frame force and a body-frame torque in one value, unlike §11.1, and it crosses the
airlock only through a typed conversion (§33.17). Six further facts found later are counted separately and recorded with their admission tests in the last subsection of §33: two cases in which the bound verifier of fs-sos returns a value that is not a bound; welded masses omitted by the suite's G1 catalog builder; a series limit that refuses the differential of the SE(3) exponential beyond 32 m of translation; a swept-sphere helper that covers straight motion only; and two more convention hazards, the sign of the legacy transformer in fs-phs and a ray direction renormalized under a scaled instance in f3d-math.
FrankenRobotics is therefore the missing product on top of a mostly present foundation. FrankenSim is the referee and the bedrock, consumed by path behind one airlock crate and never forked (§33). FrankenRobotics owns the batched Lie-algebraic rollout engine, the scenes, the sensors, the learning loop, the mission layer, the adversary, the record and the agent-facing studio.
1.8 The two machines
The program has two machines and uses each for what it is good at (DESIGN; orchestration across them is F, §21 and §36).
Linux host trj |
This Mac Studio | |
|---|---|---|
| Hardware (FACT) | Two RTX 4090, 24 GB each, no link between the cards; often heavily loaded | M5 Ultra: 36 CPU cores, an 80-core GPU, 256 GB of unified memory; a measured Metal working-set limit of about 222 GiB and a largest single buffer of about 167 GiB; memory bandwidth 1.2 TB/s by the vendor's figure |
| Role | Training and high-throughput rollouts on CUDA; the two cards run independent shards | The world bank and large scenes; proof and symbolic work; the sampled referee; replay; the viewer; large-memory jobs; a Metal rollout tier whose throughput is a measured number (BET-09) |
| Crosses the boundary | Seeds, policy weights and receipts only | Seeds, policy weights and receipts only |
Scale facts that follow from byte models and one incumbent measurement (FACT and derivation, Appendix H): per-world state is about 20–53 kB for a quadruped and 56–180 kB for a humanoid, so 8,192 quadruped worlds or 4,096 humanoid worlds use about 1 to 3% of a 24 GB card and memory does not bind at the planned batch sizes; a 2 cm home scene is tens of megabytes and a 2 cm warehouse several gigabytes, which is why the Mac holds the world bank; a CPU referee at 0.2–1.0× real time per core on 24 cores can shadow about 0.1–2.4% of GPU world-seconds, which is why referee work is sampled; the figure scales with referee speed and is measured at M2. Whether the step on an RTX 4090 is bound by memory bandwidth at these state sizes is a HYPOTHESIS for the M2 profile: bandwidth arithmetic suggests it, and an incumbent step trace with 47% of its time in the solver does not show it. On the Mac, the BET-09 threshold needs about a tenth of the available memory bandwidth, so command-buffer overhead and the GPU's arithmetic rate are what must be measured.