SmolVLA HSR V1 — Fine-tuned for Mobile Manipulation
Model Details
- Base model: lerobot/smolvla_base (450M params)
- Trainable params: 99.9M (expert only)
- Training: 50K steps, 2×H100 PCIe 80GB, ~13.3h
- Final loss: 0.0070
- Dataset: 388K episodes from AIRoA ICRA 2026 Mobile Manipulation Challenge
State/Action Space
- State: obs_state[8] + ee_pose[6] + wrench[6] = 20 dims → padded to 32
- Action: action.relative[11] → padded to 32
- Action chunk: 50 steps
- Cameras: head (480×640) + hand (480×640)
Training Recipe
- Optimizer: AdamW (lr=2e-4, cosine decay to 5e-6, warmup=500)
- Batch size: 128 per GPU (256 total)
- SDPA monkey-patch for 3× attention speedup
- DDP on 2 GPUs
Usage
from lerobot.common.policies.smolvla.modeling_smolvla import SmolVLAPolicy
policy = SmolVLAPolicy.from_pretrained("ashley77/airoa-smolvla-hsr-v1")
