SmolVLA — rm65b-sort-v0
LeRobot SmolVLA policy fine-tuned on rm65b-sort-v0 (bimanual RM65 sort task,
3 cameras, 198 episodes / 91.5k frames @ 30 Hz, 1 task).
Three candidate deployment checkpoints from the same training run.
| Subfolder | Train step | Final train loss |
|---|---|---|
checkpoint-025000 |
25,000 | 0.036 |
checkpoint-035000 |
35,000 (best logged loss) | 0.033 |
checkpoint-045000 |
45,000 (last on disk) | 0.034 |
Loss flatlined after step ~30K once the cosine LR clamped at its minimum (2.5e-6); the three checkpoints sample the early plateau, best loss, and final state respectively.
Training summary
- Architecture: SmolVLA (LeRobot)
- Batch size 128, cosine LR (peak ~1e-4, min 2.5e-6), AMP
- Hardware: 1× A100-80g, ~24 h wall-clock
See each checkpoint-XXXXXX/train_config.json for the full reproducible config.
Usage
from huggingface_hub import snapshot_download
from lerobot.policies.smolvla import SmolVLAPolicy
ckpt_dir = snapshot_download("JayCao99/smolvla-rm65b-sort-v0", allow_patterns="checkpoint-045000/*")
policy = SmolVLAPolicy.from_pretrained(f"{ckpt_dir}/checkpoint-045000")