FrankenRobotics

Robotics lab for AI agentsIn design

Name the job. Get every robot that can do it.

FrankenRobotics is a robotics lab that AI agents run. Give it a mission and a parts catalog. It designs the bodies, works out the controllers, tests them in thousands of worlds at once, and hands back every design that passes, with a build packet for the one you pick.

Pick a place

0 of 99 fit

0 too wide0 too tall0 too long

Drag a slider. Robots that still fit stay lit. FrankenRobotics checks size first, then torque, grip, sensing and battery.

The lab

From a mission to a machine.

Most robot projects start with a body and ask what it can do. FrankenRobotics starts with the job and works back to every body that could do it.

What it does

Six tools that share one model.

Simulation, design, search, testing, calibration and proof all read the same description of the robot, so whatever one tool learns, the others can use.

  • Simulate

    Thousands of worlds at once

    Legged, wheeled, tracked and hybrid robots run side by side on the GPU, across generated terrain and through whole generated buildings.

    ArenaCortex

  • Design

    A studio agents can drive

    One command line and one MCP server to create a robot, check it, run it, compare versions and see what an edit changed.

    AtelierTether

  • Ask

    The whole family in one query

    Give it a mission and a parts catalog. Get every design that works, the reason each of the others fails, and the cheapest test for the close calls.

    Augur

  • Break

    An opponent for every robot

    Nemesis hunts for the world that beats your design and strips it to the details that matter. Autopsy replays the fall until the cause is clear.

    NemesisAutopsy

  • Build

    A path to the workshop

    Calibrate against real measurements, then export a build packet with parts, solid geometry, tolerances and fatigue margins.

    GroundFoundry

  • Prove

    Math you can check

    One Lie-group model produces the GPU kernels, the tests that police them, and certificates that a separate program re-checks offline.

    MarrowLemmaSpark

Watch an agent run it

Twenty commands to a qualified robot.

The mission: carry a load across a wet strip of kitchen floor. The agent finds a cheap design, breaks it, fixes the wheel and puts it through the exam.

How agents drive it

Step through it, or press play

Worked example

agent-17

$ frob session open --scopes design:write,mission:run,nemesis:run,autopsy:run,sim:run,qual:run --budget gpu_s=1800,cpu_core_s=7200 --json

exit 0: session s:2f6a…, namespace agent-17

$ frob mission new carrier-wet-strip --template delivery-and-placement --json

exit 0: query q:5b1e… (.fmission)

$ frob mission check q:5b1e… --json

exit 0: domain nonempty; ∃ policy per design, ∀ worlds, ∀ executions; 2 topologies, 412 cells

$ frob mission synthesize q:5b1e… --estimate --json

640 GPU-seconds, 90% interval 512 to 800

$ frob mission synthesize q:5b1e… --budget gpu_s=900 --key syn-1 --sync-wait-ms 5000 --json

exit 8: job j:77c0…, 0 of 412 cells

$ frob job wait j:77c0… --json

exit 2: inner 61 cells; outer 233 of 412; 179 excluded with witnesses; 172 unresolved

$ frob mission show q:5b1e… --set inner --order cost --limit 3 --json

cheapest inner cell c:0007, part cost 1,480

$ frob mission show q:5b1e… --set unresolved --order cost --limit 3 --json

cheapest unresolved cell c:0204, cost 1,120; next_test: attack c:0204's best controller

$ frob mission refute q:5b1e… --cell c:0119 --json

excluded for every controller: the loaded width cannot pass the door; smallest relaxing change: track −14 mm

$ frob design new carrier-b --from-cell q:5b1e…/c:0204 --json

design d:a3f1… at the cell centre

$ frob nemesis search carrier-b --mission q:5b1e… --budget gpu_s=240 --key nem-1 --json

exit 2: 3 failing worlds inside the family

$ frob nemesis shrink w:9d04… --key shr-1 --json

exit 2: .fmcw, 2 edits from nominal: counter lip 18 → 22 mm; wet strip 40 mm before it

$ frob autopsy intervene r:5c11… --key aut-1 --json

exit 2: holding strip friction at its dry value removes the failure in 60 of 64 forks; the placebo flips 2 of 64; stable

$ frob design edit carrier-b set-param wheel.radius 0.055 0.065 --key ed-1 --json

new version d:b77e…

$ frob result revalidate carrier-b --since d:a3f1… --json

14 of 19 prior results inherited; 5 need rerun, 41 GPU-seconds

$ frob nemesis search carrier-b --mission q:5b1e… --budget gpu_s=240 --key nem-2 --json

exit 0: none in 4,096 worlds

$ frob eval qualify carrier-b --mission q:5b1e… --dry --json

submission 1 of this lane: α₁ = 0.025; 200 sealed worlds; qualified is reachable only with at most 3 failures

$ frob eval qualify carrier-b --mission q:5b1e… --key qual-1 --json

exit 0: qualified, 1 failure of 200; receipt qr:e2a0…; no per-world data returned

$ frob prove bundle carrier-b --mission q:5b1e… --receipt qr:e2a0… --out carrier-b.frobcert --json

$ frob-verify carrier-b.frobcert

offline: bundle checks

Step 1 of 20

Open a session

The agent gets a budget of GPU and CPU seconds, and a token that can’t open the sealed test worlds.

Where the time goes

Training eats the clock.

In a typical design loop, 85 to 90% of the time goes into training a controller. FrankenRobotics rules out the bodies that can’t work first, then trains the few that might.

One pass of the loop: edit, compile, train, evaluate

Robot dogabout 6 minutes
Training, about 85%The rest
Humanoidabout 31 minutes
Training, over 90%

80 to 1,600candidates trained in one published design campaign

7.5 to 30GPU-minutes to train each of them

15.8%turn out to be duds after the full training run

2 secondsis the budget for the first six FrankenRobotics checks

Times are composed from published training runs.

What changes for you

Twelve jobs it’s built for.

Everyday problems in robot design, how they go today, and how they go with FrankenRobotics.

  1. 01Catch weak knees before training
  2. 02Go back to one world out of 4,096
  3. 03Get the family, skip the guessing
  4. 04Run 256 different robots together
  5. 05Decide 300 designs in an hour
  6. 06Shrink a failure to three details
  7. 07Name the cause of a fall
  8. 08Stop lucky designs getting through
  9. 09Fix the room for $12
  10. 10Walk a new body straight away
  11. 11Test in your own kitchen
  12. 12Hand the winner to a workshop

Catch weak knees before training

Today

An agent edits the robot’s XML, starts training, and learns up to an hour later that the knees can’t carry it.

With FrankenRobotics

frob design check answers in under two seconds. The knees are 31% short on torque, and three repairs are ready to apply.

Read the full scenario

Go back to one world out of 4,096

Today

World 1,337 fills with NaNs. To look at it, you re-run the whole batch on the same GPU.

With FrankenRobotics

frob sim replay --world 1337 rebuilds that one world on the CPU from its address, stops at the first bad step and shows the contact that started it.

Read the full scenario

Get the family, skip the guessing

Today

An agent scores candidates one at a time and reports the best. Nobody knows about the bodies it never tried.

With FrankenRobotics

One query returns every stance that carries the load and proves the rest can’t. The limit is exact: 3/14 of a meter from centerline to foot.

Read the full scenario

Run 256 different robots together

Today

A varied population often means a rebuild for each design, with compiles that can take minutes.

With FrankenRobotics

One batch steps 256 bodies with three different leg counts, 32 worlds each, on a single RTX 4090. Nothing is recompiled.

Read the full scenario

Decide 300 designs in an hour

Today

The usual loop tries one candidate per turn.

With FrankenRobotics

An agent runs thirty edit-and-test cycles in an hour on one workstation, ten variants at a time. frob design bisect finds the edit that broke stair climbing.

Read the full scenario

Shrink a failure to three details

Today

Random testing says 0.4% of runs fail, and the agent reads logs to guess why.

With FrankenRobotics

Nemesis finds a failing world and strips it down: a door sill 9 mm higher, one slippery tile, a depth frame 40 ms late. Take away any one and the robot makes it.

Read the full scenario

Name the cause of a fall

Today

An agent sorts the anomalies by time and guesses.

With FrankenRobotics

Autopsy rewinds to just before the fall and changes one thing at a time. Lifting the knee torque limit saves 94% of the replays; better grip saves 8%. It was the knee.

Read the full scenario

Stop lucky designs getting through

Today

Submit enough candidates to the same test and one passes by chance.

With FrankenRobotics

Every submission meets fresh sealed worlds, and one error budget covers the whole campaign, even after 500 retries.

Read the full scenario

Fix the room for $12

Today

Every repair an agent considers is a change to the robot.

With FrankenRobotics

The clearance map finds the culprit, a 19 mm threshold. A $12 rubber ramp fixes it with the robot untouched. A bigger drive wheel costs $140. You see both prices.

Read the full scenario

Walk a new body straight away

Today

Every new design waits for its own training run: about seven minutes for a quadruped, up to an hour for a humanoid.

With FrankenRobotics

Lengthen the femur 8% and widen the stance, and the policy trained for that family of bodies walks the new one, at 78% of a specialist’s score.

Read the full scenario

Test in your own kitchen

Today

You test in a home that somebody else modeled.

With FrankenRobotics

A walkthrough video of your kitchen becomes a private test scene. Candidate robots work in it, and the scan never leaves your machine.

Read the full scenario

Hand the winner to a workshop

Today

The loop ends with a score.

With FrankenRobotics

Foundry turns the design into a parts list, solid geometry, a tolerance stack that closes and fatigue margins from the simulated loads, with one link 18% lighter at the same margin.

Read the full scenario
1 of 12

The body

Legs, wheels, tracks, or all three.

A robot is a set of modules with typed ports. Swap a motor, stretch a femur or bolt a wheel onto a leg, and the new body is ready to simulate with nothing to recompile.

Starting points

Eighteen bodies to start from.

A robot dog, a humanoid, a Mars-style rover, a snake. Each one is built from the table beside it, so stretching a link or moving a hip gives you a new machine.

How a robot is described

Pick a body, then drag to turn it

Interactive

Drag to turn

Motors
12
Symmetry
C₂ × C₂
Links

Body 1 of 18

Robot dog

Four legs, three joints each.

Trots, paces and bounds. On stairs its knees run hot, and the studio sees it coming.

linkparentjointoffset, m

The design check

Catch it in two seconds.

Eight checks run before any simulation, and the first six take two seconds between them. When one fails, the answer comes with a repair the agent can apply.

All eight checks

Hand it a faulty robot, then apply the repair

Worked example
  1. 1
    FormUnits, ports and parts agree
    100 ms
  2. 2
    MassWeight and inertia add up
    50 ms
  3. 3
    MechanismJoints close and links don’t collide
    400 ms
  4. 4
    MotionWhich ways it can really move
    200 ms
  5. 5
    StructureSymmetry and power balance
    150 ms
  6. 6
    SizingTorque, speed, heat, tip-over, battery
    600 ms
  7. 7
    PlaceReach, sightlines and clearance
    2 s
  8. 8
    InformationSensors can tell the worlds apart
    10 s

frob design check

One batch

256 robots, all different.

Four, six or eight legs, every limb its own length, all stepping together on one GPU. A new body is one more row in a table.

How mixed batches work

Point at a robot to read its row

Interactive
Bodies
256
Links moving
0
Recompiled
0

The world

Real rooms, real trouble.

Robots get stuck on thresholds, slide on wet mats and jam in corners four centimeters too tight. FrankenRobotics generates whole buildings full of those hazards and knows the size of every one.

Generated buildings

Seven kinds of building, a new plan every seed.

Homes, apartments, offices, warehouses, factories, shops and commercial kitchens, each with its own rooms, clutter and traps.

How places are generated

Pick a place, then try another seed

Interactive

Hazards in this one

    What it tests

    Clearance

    Will it make the turn?

    Quick checks get this wrong in both directions. The clearance map looks at every position and heading, then hands back a path, or proof that none exists.

    How clearance is decided

    Stretch the robot until it jams

    Interactive
    • Clear
    • Too close to call
    • Blocked

    Quick check, optimistic

    Quick check, cautious

    Clearance map

    Exact answer

    The exact answer for this corner is 2√2·a − 2b, for a corridor a wide and a body b wide.

    Thresholds

    Bigger wheel, or bigger motor?

    A bigger wheel needs less grip to climb a threshold and more torque to do it. At one setting only the large wheel gets over, and at another only the small one does.

    Wheels, tracks and traction

    Change the grip and the torque

    Interactive
    WheelLeverTorqueGrip needed

    The answer

    One question, every answer.

    Ask which robots can carry a tray across the kitchen and you get the whole family: the designs that work, the ones that can’t and why, and the handful worth testing.

    A family, live

    14,400 runners, released together.

    Each pixel is a different one-legged robot, let go at 5 m/s. Most fall within a few steps. The ones still running trace the bright crescent, and that crescent is the answer.

    Move the window to watch its robots run

    Time
    0.0 s
    Still running
    14,400
    Window
    68°, 34 kN/m
    Leg stiffness, 5 to 55 kN/m
    56°Landing angle80°
    Still running, soft leg to stiffFallen, late to early
    The spring-legged runner of Seyfarth, Geyer, Günther and Blickhan (2002), computed in your browser.

    The family map

    See every design at once.

    Each point on this map is a four-legged carrier with a different stance. Green boxes work, red ones can’t, and amber ones straddle the edge. Refine them and the amber shrinks to a line.

    How missions are answered

    Press refine, then point at any box to see why

    Interactive

    Works0%

    Too close to call100%

    Can’t work0%

    Point at a box

    The reason it works, or doesn’t, shows up here.

    An exact limit

    How narrow can it get?

    This robot has to balance a 40 kg load anywhere in the marked rectangle. Squeeze it until a foot gives out. The limit is 3/14 of a meter from centerline to foot, and you can check the fraction by hand.

    The worked answer

    Squeeze the robot and drag the load

    Interactive

    Trade-offs

    Better sensor, or stronger motor?

    A cart has three pushes to stop on the mark, and it doesn’t know how strong its own motor is. One sensor reading can tell it, if it arrives in time. Trade one part against another and watch the answer flip.

    Sensing, timing and control

    Make the sensor worse, then make the motor stronger

    Interactive

    The crash lab

    Break it here first.

    FrankenRobotics hunts for the world that beats your robot, boils it down to the details that matter, and replays the failure until the cause is clear.

    Nemesis

    The smallest world that breaks it.

    This robot fails in a world with five things wrong. The search takes them away, test by test, until it finds the pair that’s enough on its own.

    How Nemesis attacks

    Press run and watch it narrow down

    Interactive

    Autopsy

    Find out why it fell.

    Go back to just before the fall, change one thing and replay it 200 times. Then make a change that shouldn’t matter, and subtract whatever that saves.

    How Autopsy finds a cause

    Change how many replays each fix saves

    Worked example

    The sealed exam

    Luck doesn’t pass.

    Let an agent resubmit worthless designs 500 times and an ordinary 5% test will pass one sooner or later. FrankenRobotics spreads a single error budget over every retry.

    How qualification works

    Submit 500 worthless designs, 400 times over

    Interactive

    Ordinary 5% test

    0

    campaigns fooled

    FrankenRobotics

    0

    campaigns fooled

    Submission 0 of 500

    The engine

    Algebra all the way down.

    Rust at the core. One set of kernels for CUDA, Metal and the CPU. A slower, stricter physics engine as referee. Underneath it all, forty pieces of mathematics that each have a job.

    Twenty-three parts

    Assembled, like the monster.

    Every subsystem is named for what it does to the robot. Point at a part to meet it.

    Point at a name, or at the machine

    • Cortex

      The gym. Trains one policy for a whole family of bodies.

    • Sensorium

      Cameras, lidar, touch, and the clock that stamps every reading.

    • Reflex

      Balance, gaits and a controller worked out from the model.

    • Spark

      The GPU kernels for CUDA and Metal, with a CPU twin.

    • Sinew

      Motors, gearboxes, batteries, and how hot they get.

    • Skeleton

      The design language: modules that snap together.

    • Marrow

      The algebra underneath everything else.

    • Kinesis

      Motion and contact, at three levels of detail.

    • Traction

      Wheels, tracks, feet, friction and soft ground.

    • Habitat

      Terrain and buildings, generated or scanned.

    • Airlock

      The gateway to FrankenSim’s physics.

    • Phylum

      Families of different bodies in one batch.

    • Augur

      Answers missions with every design that works.

    • Atelier

      The studio: edit, compare, trace, repeat.

    • Nemesis

      Hunts for the worlds that beat a design.

    • Autopsy

      Replays a failure until it finds the cause.

    • Chronicle

      Records every run, joule and result.

    • Lemma

      Proofs, and an offline program that checks them.

    • Gauntlet

      The tests that the lab itself has to pass.

    • Foundry

      Parts, tolerances, fatigue life and build packets.

    • Ground

      Tunes the model to measurements of real robots.

    • Tether

      The command line, MCP server and viewer.

    • Arena

      Tasks, scoring and the sealed exam.

    Symmetry

    Teach one leg, get the rest.

    A robot dog’s legs mirror each other, so a policy can share what it learns. Bolt on a lidar or an off-center load and some of that symmetry goes. FrankenRobotics works out what’s left.

    Two kinds of symmetry

    Mount something and watch the symmetry change

    Interactive
    The mechanismC₂ × C₂
    With sensors and loadC₂

    Worked by hand

    Check its arithmetic.

    Six problems small enough for paper. Each shows the right answer beside the mistake it catches.

    The whole answer book

    Pick a problem

    Interactive

    Roadmap

    What ships first.

    The design is public today. The build follows in this order, so there’s something to use long before the end.

    Every milestone
    1. Today

      The design

      The full plan, public and searchable.

    2. Step 1

      Foundations

      One kernel running on the CPU, CUDA and Metal.

    3. Step 2

      The CPU studio

      Design checks, the referee and the first mission answers, with no GPU needed.

    4. Step 3

      GPU batches

      256 different robots at once, on CUDA and on Apple silicon.

    5. Step 4

      Worlds and adversaries

      Generated buildings, Nemesis, Autopsy and the sealed exam.

    6. Step 5

      Families at scale

      Shared policies, co-design and big mission queries.

    7. Step 6

      The workshop

      Calibration against real robots, certificates and build packets.

    8. Step 7

      The ten-times test

      Sealed missions, head to head with the best stack anyone can assemble.

    In design, in the open

    Follow the build.

    FrankenRobotics is being designed in public. The whole plan is here: 70 sections, about 643,000 words and every equation, free to read and easy to search. Watch the repository to catch the first release.

    You design robots

    Bodies, buildings, missions and the workshop.

    SkeletonHabitatAugurFoundry

    You build simulators

    Dynamics, contact, GPU kernels and performance.

    KinesisTractionSparkSpeed

    You build agents

    Commands, budgets and the sealed exam.

    AtelierTetherArenaSafety

    You’re here for the math

    Lie groups, cones, certificates and proofs.

    ConstructionsMarrowLemmaCompendium