Adversarial scene variation
Perturb object pose and contact conditions across parallel task instances.
RTX WORKER ONLINE · ISAAC LAB 2.3.2
Stress-test robot intelligence before reality does.
BreakBot probes manipulation policies against physics shifts, control degradation and randomized edge cases—then returns measurable evidence before hardware is at risk.
32 seeds
rollouts per run
0—100
comparable score
≈75 sec
typical runtime
Franka
live task
The stress-test terminal
From one clean lift to the full failure surface.
A production-shaped queue sends versioned jobs to a private RTX worker. Isaac Lab executes deterministic rollouts, renders the evidence, and returns metrics without exposing worker credentials.
Physics perturbation
Mass, friction, pose and actuator limits
32-seed evaluation
Repeatable rollouts expose brittle behavior
Controller variants
Joint position, absolute IK and relative IK
Failure forensics
Robustness, collisions, timing and replay
RTX execution
Native NVIDIA Isaac Lab worker on private compute
Rendered evidence
Shareable H.264 output attached to every run
NOMINAL ENVELOPE
When the policy holds.
Normal friction. Full actuator authority. Low latency. The Franka policy secures the cube, clears the table and completes the trajectory across the expected operating range.
MASS
0.35 kg
MOTOR
92%
LATENCY
18 ms
FAILURE DISCOVERED
When the cube drops.
Increase payload, cut friction, weaken the motors or delay control. BreakBot maps the boundary where a safe-looking policy becomes a hardware incident.
Find the breaking pointExtensible test coverage
Built to interrogate the physical stack.
BreakBot currently exposes one small model for public tests. The worker contract is designed for richer manipulation benchmarks when more compute is available.
Precision manipulation
Probe grasp tolerance, small-object control and placement sensitivity.
Massively parallel policy trials
Turn randomized failures into a reproducible robustness surface.
Controller boundary analysis
Compare joint-space and inverse-kinematics behavior near operating limits.
Cross-platform evaluation
A worker-ready contract for additional small manipulation models.
Small models. Real compute. Inspectable results.
Public tests run on privately operated GPU capacity and may wait in a queue. No model upload is required or accepted in this version.
The intelligence assurance layer
Every robot policy will need an adversary.
Autonomy is moving from controlled demos into factories, hospitals and infrastructure. BreakBot is building the verification terminal between a trained policy and an expensive machine: configure the stress, execute at scale, preserve the evidence, ship with confidence.
01
Probe
Attack the operating envelope.
02
Measure
Quantify robustness across seeds.
03
Replay
Inspect physics-derived evidence.
04
Improve
Turn failures into better policies.
Today: a public Franka benchmark on private RTX compute. Next: more policies, more task families and deeper adversarial evaluation—without changing the submission or evidence architecture.