Inoki Labs

Crash testing for robot policies

BreakPoint runs a trained walking policy through thousands of simulated conditions and shows you which ones make it fall.

Talk to us

Normal conditions

Same policy, same seed

Keeps walking

  1. Normal conditions, 0.00 seconds0.00 s
  2. Normal conditions, 0.56 seconds0.56 s
  3. Normal conditions, 0.76 seconds0.76 s
  4. Normal conditions, 0.86 seconds0.86 s

Found by BreakPoint

40 ms latency · 13% sensor dropout · 4° slope

Falls in 0.86 s

  1. Found by BreakPoint, 0.00 seconds0.00 s
  2. Found by BreakPoint, 0.56 seconds0.56 s
  3. Found by BreakPoint, 0.76 seconds0.76 s
  4. Found by BreakPoint, 0.86 seconds: the robot has fallen0.86 s
Falls in 56 of 64 simulated runs under these conditions, and 0 of 64 in normal ones.

How it works

  1. 01

    Search

    Changes friction, mass, latency, sensor noise and more across thousands of runs at once on a GPU.

  2. 02

    Verify

    Replays each fall exactly, then runs it again in a second physics engine.

  3. 03

    Gate

    Compares every new checkpoint with the current one in CI and fails the build if it got worse.

Results so far (simulation)

2×

as many falls found in realistic conditions as random sampling, same compute

1%

false alarm rate for the CI check. A simple any-regression rule fails 50% of unchanged builds

~50%

fewer falls than standard domain randomization after retraining, on Go1 and Spot

  • Go1
  • Spot
  • Barkour
  • G1
  • T1
Go1SpotBarkourG1T1

Team

Vihan Aggarwal

CEO and CGO

MIT CSAIL · IEEE Best Paper, AIIoT 2026

Subhi Stephan

CTO and CMO

Duke BME and ECE · MIT research

Backed byAntler

Contact

If you train policies for legged robots, we'd like to run BreakPoint on one of yours.