Thali multi-task SmolVLA
lerobot/smolvla_base fine-tuned on Prashant-77/thali_all (1050 scripted-expert episodes,
7 skills, language-conditioned, 3 cameras). Camera keys are renamed at train and inference time: overhead → camera1, wrist_a → camera2, wrist_b → camera3
(--rename_map; runtime/executors.py applies the same map).
Checkpoints in this repo. Root = 20 000 total steps. step_14000/ = the best per-skill checkpoint (14 000 steps at batch 16 on a T4; the last
6 000 steps ran at batch 4 with a fresh optimizer on a smaller GPU and lost ground).
Policy-only success per skill from task-consistent start states, 20 held-out seeds (eval/skill_eval.py --kind smolvla):
| skill | step 14 000 | step 20 000 |
|---|---|---|
| open_drawer | 14/20 | 11/20 |
| pick_place_fork | 0/20 | 0/20 |
| pick_place_plate | 2/20 | 0/20 |
| pick_place_mug | 0/20 | 0/20 |
| handoff_spoon | 0/20 | 0/20 |
| hold_mug | 8/20 | 3/20 |
| pour | 0/20 | 0/20 |
The scripted expert reaches 9/10 on the full task; the per-skill ACT baselines at 50k steps reach drawer 20/20, hold 18/20, plate 15/20, fork 14/20. Repository, evaluation scripts and every results file: https://github.com/Prashant-thakur77/THALI
- Downloads last month
- 12
Model tree for Prashant-77/thali_smolvla
Base model
HuggingFaceTB/SmolLM2-360M