LaundryBot
Two SO-101 arms in an ALOHA-style setup, fine-tuning open models (LeRobot, SmolVLA, pi0-family) instead of training from scratch.
How it learns
Clips Simulation
Scripted two-fold
Newton VBD cloth sim, 30 cm towel, trial 0 of 20. Scripted motions, not a learned policy.
laundrybot/rounds.md:41
Milestones
| 1 | Fold a hand towel twice: at least 60% success over 20 filmed trials, scored by a written checklist. | Not started on hardware |
| 2 | Fold a T-shirt: 50% or better. | Not started |
Latest offline result
2026-10-23: a small action-chunking policy (state only, no cameras) trained on four pooled public towel datasets versus one dataset alone. Scored by action error over 10 and 25 future steps, lower is better.
| Arm | 10-step error | 25-step error |
|---|---|---|
| Single dataset | 0.1321 | 0.2039 |
| Pooled (4 datasets) | 0.1442 | 0.2223 |
| Baseline: hold current position | 0.1744 | 0.3374 |
Pooling lost this round. Both learned policies beat the hold-position baseline. Suspected causes, still unverified: calibration offsets between datasets and one dataset with much slower folds. Per-dataset normalization is being tested next. These are offline metrics on recorded data, not robot success rates.