Spa-Bench physical rollouts — pi0.5
Physical SO-101 rollout recordings for the pi0.5 Spa-Bench checkpoint at epoch 12, step 76,596.
Scope and counts
The scored benchmark protocol is episodes 0–899: 300 primary withheld/OOD spatial trials and 600 diagnostic/control trials. Their prompt sequence matches the MolmoAct2 and GR00T-N1.7 Frozen LLM full releases exactly.
This repository contains 920 episodes, 781,538 frames, 356 unique instruction strings, two 480×640 RGB streams, and six-dimensional state/action at 30 FPS. Episodes 900–919 are additional Size Recognition recordings using six familiar smallest/largest prompt forms. They are retained as raw research records but are excluded from the thesis's 900-rollout analysis.
Reported physical results
| Condition | Successes | Rate |
|---|---|---|
| Familiar/in-distribution | 82/120 | 68.3% |
| All withheld/OOD spatial configurations | 122/300 | 40.7% |
| Matched OOD subset | 54/120 | 45.0% |
| Matched direct-manipulation controls | 100/120 | 83.3% |
Outcome labels and analysis groupings are maintained in the Spa-Bench evaluation spreadsheets, not as a frame-level success field in this LeRobot dataset. Binary success uses the task-specific criteria defined in the report; no partial credit is assigned.
Structure and use
Each trajectory stores the instruction/task index, six absolute SO-101 joint positions for state and action, and synchronized middle/wrist video. This is an evaluation record, not a fine-tuning split. It is intended for audit, replay, qualitative failure analysis, and benchmark reproduction.
The results apply only to the linked checkpoint, workspace, object layouts, and physical protocol. Do not use the trajectories to inflate the training set and then report performance on the same benchmark as held out. Robot data can enable unsafe behaviour; deployment requires safeguards and human supervision.
Citation
Please cite the completed Spa-Bench MSc report and the thesis artifact.
- Downloads last month
- 685