arrow-hf/smolvla-robotwin-place-object-basket-50ep

SmolVLA RoboTwin place_object_basket (50 ep, single instruction)

SmolVLA policy fine-tuned on 50 demonstration episodes of the place_object_basket task from RoboTwin 2.0 (demo_clean config), built on the SmolVLA-RoboTwin pretrained base (lerobot/smolvla_robotwin).

See also the multi-instruction counterpart: arrow-hf/smolvla-robotwin-place-object-basket-50ep-multi

Task & Training

• Robot: Agilex dual-arm, end-effector control (16D state, 16D action)
• Cameras: 3 RGB streams (240×320, D435)
• Instruction mode: single fixed instruction (Strategy A)
• Training: bs=32, 6000 steps (~10-25 epochs), AdamW lr=1e-4, cosine warmup=300/decay=6000
• Chunk size: 50

Evaluation

RoboTwin 2.0 sim (demo_clean), 10 episodes, max_steps=400, action_chunk_exec=50, eval instruction "place the object in the basket".

Success rate: 3/10 (30%)

Surprising finding: On this task, the single-instruction version actually underperforms the multi version (30% vs 50%) — see the multi counterpart for full discussion.

Usage

from lerobot.policies.smolvla import SmolVLAPolicy
policy = SmolVLAPolicy.from_pretrained("arrow-hf/smolvla-robotwin-place-object-basket-50ep")

At inference, use action_chunk_exec=50 (full chunk).