RECORDED EVIDENCE VIEWER
Did the robot carry it,
or just land near the target?
Inspect a scripted SO-101 manipulation trial, the checks behind it, and the failures we kept.
Partial prototype. This page replays saved results. The robot is controlled by scripted IK; there is no trained task policy, VLM planner or completed dinner-table task.
01 / Watch the actual run
MuJoCo simulation · left arm active102 seconds · silent, with on-screen text. The 19-second simulation recording starts at 0:32. No physical robot or model-driven action is shown.
02 / Inspect all ten variations
6 passed / 4 failedCube position changes only: XY ±5 mm and yaw ±0.12 radians. These are small scripted perturbations, not full-task tests or evidence of learned generalization.
Loading the committed evidence.
- Final XY error
- Minimum sustained lift
Final position alone does not determine the outcome.
| Seed | Result | XY error |
|---|
03 / A failure improved the scorer
Recorded counterexample retainedAn earlier endpoint-only check passed a cube that slipped during transfer but landed near its target. Scorer v3 rejects that trace. The regression suite checks the real failure and the corrected nominal carry.
Checks are sampled every 50 ms: both jaw contacts during hold/carry, sufficient lift, supported lowering, and stable release within 25 mm of the target. Events between samples can be missed; this is not continuous formal verification.
04 / Intel runtime preflight
Intel i5-1245U CPU · four threads · FP3242.649 ms
OpenVINO median across five calls after one warmup.
Native PyTorch median: 60.634 ms.
Compatibility, not learned behavior
ACT used random weights. No model action was executed. Numerical parity was checked on two fixtures; maximum absolute error was 0.000001386. The small timing sample does not establish a production speedup.
Inspect runtime evidence ↗05 / Review and reproduce
13 repository tests passedNo task-trained checkpoint, coordinated table-setting sequence, Core Ultra test or physical robot test is delivered. The next gate is verified demonstration collection, bounded policy training and closed-loop evaluation.