SYSTEM DOCUMENTATION · V1

Robot intelligence deserves a hostile test environment.

BreakBot is an adversarial evaluation terminal for robot policies. It converts a browser configuration into a durable GPU job, executes deterministic Isaac Lab trials, and returns evidence that can be inspected before a policy reaches expensive hardware.

Run your first test View source

Job lifecycle

One request. Four observable states.

Runs survive refreshes and retain the same shareable URL. Worker leases, heartbeats, retries and idempotent completion protect the pipeline when hardware or network conditions change.

Queued

The request is committed to durable storage and receives a stable run URL.

Running

The RTX worker atomically claims a lease and evaluates 32 deterministic seeds.

Rendering

Omniverse RTX records a native H.264 replay from the official task scene.

Completed

Metrics, provenance and replay become permanent, refresh-safe evidence.

Parameter surface

Push the policy beyond the clean demo.

The terminal exposes physical, sensing and control stressors. Color zones communicate the intended operating envelope; submitted numeric values are stored exactly with the run.

ControlEffectRange

Forward reachCube distance from the robot base0.30–0.70 m

Lateral offsetOff-axis displacement of the grasp target−0.30–0.30 m

Payload massPhysX rigid-body load0.05–30 kg

Contact frictionCube, gripper and table coefficient0–2 μ

Motor authorityJoint effort and velocity multiplier20–120%

Sensor noiseGaussian observation perturbation0–0.25 σ

Control latencyDelay injected before commands0–250 ms

Random seedReproducible perturbation family0–999999

Evidence contract

A result is more than a video.

Every completed run returns the evidence required to reproduce, compare and diagnose policy behavior.

Robustness score

Comparable 0–100 aggregate across all rollouts.

Seed matrix

Pass/fail evidence and score for each of 32 seeds.

Physical events

Drops, collisions and terminal task failures.

Completion time

Measured execution timing per deterministic seed.

RTX replay

Public H.264 artifact with a verified content hash.

Provenance

Simulator, Isaac version and execution GPU.

System architecture

The website never touches simulator internals.

A versioned contract separates the public interface from GPU execution. The same API can later dispatch additional Isaac Lab executors without changing run URLs or result pages.

WEB

Validate + submit

QUEUE

Persist + lease

RTX

Simulate + render

RESULT

Metrics + MP4

POST /api/runs
GET  /api/runs/:id
GET  /api/queue?runId=:id

contract: 2026-09-03-franka-v1
seeds:    32
renderer: Omniverse RTX
artifact: video/mp4

Compute policy

Real simulation has a real cost.

Public jobs use privately operated RTX compute and are processed in queue order.

The public terminal is limited to a small Franka model; complex systems require much longer GPU runs.

No policy files, credentials, payments or personal data are accepted by the MVP.

Displayed completion times are estimates and can change with startup, rendering and queue load.

Ready to find the boundary?

Start with the nominal profile, move into the amber zone, then deliberately target a red failure configuration. Every run is isolated behind its own permanent result URL.

Open stress-test terminal