Predict
Forecast body, contact and action response across multiple horizons.
Research & demonstrations
Repair the model while the machine is still operating.
From a single robot to a coordinated production cell, we study how a model can identify a changed mechanism and support a verified recovery.
Research question
Can a learned model detect a hidden mechanical change, produce a bounded correction and recover the active mission without broad retraining?
The demonstration
One abrupt fault. One continuous recovery.
The command response of the left leg reverses after impact. A08 detects and localizes the change, verifies a local sign correction and reaches the original target during the same episode.
0.22 sdetect
0.03 scompile
4 / 4recovered
0resets
Production line · Simulation
One robot slows. The line adapts.
Two arms, one conveyor. When the packing robot loses 30% of its motion response, A08 repairs the line model and adjusts the conveyor and first robot's timing.
3 / 5Original schedule
5 / 5A08 repair + replan
The repair loop
One falsifiable sequence.
Detect
Identify when observed mechanics no longer match the current model.
Diagnose
Use a bounded exposure to locate the affected mechanism.
Verify
Test the candidate correction on evidence it did not use to fit.
Continue
Route the accepted model and resume the same mission.
Unitree G1 evidence
The claim is narrower than the ambition.
The G1 confirmation and command-inversion demonstration use different workloads. The production-line case is a separate deterministic demonstration, not part of these results.
Infrastructure
Eight isolated qualification runs completed across four NVIDIA L4 GPUs with exact same-stack replay.
Controller
The frozen public controller completed 99 of 100 sealed reach, disturbance and stable-stop missions.
Self-model
Three independent 12.39M-parameter members passed every validation and once-only nominal audit gate.
Change detection
The frozen detector found 64 of 64 changed worlds by observation 100 with zero nominal false positives.
Verified actuator-loss repair
One typed parameter restored 48 of 48 fresh missions from zero, with no repaired violation.
Inverted-leg recovery
A categorical local correction recovered 4 of 4 fresh runs versus 0 of 4 frozen, with no reset.
Binary outcomes
A partial goal score is never presented as mission success.
Independent verification
The model cannot authorize its own behavioral update.
Simulator evidence
No physical safety, sim-to-real or Unitree endorsement claim is implied.
Physical validation
Help us take the next experiment onto real hardware.
We are looking for a robotics partner with a programmable platform, telemetry access and a repeatable mission.