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 single-instruction counterpart: arrow-hf/smolvla-robotwin-place-object-basket-50ep

Task & Training

Robot: Agilex dual-arm, end-effector control (16D state, 16D action)
Cameras: 3 RGB streams (240×320, D435)
Instruction mode: per-episode random instruction from RoboTwin's 100 variations (seed=42)
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: 5/10 (50%)

Surprising finding: This is the only task in our 8-task benchmark where multi-instruction training outperforms single-instruction (50% vs 30%, +20pp). Likely because the task involves multiple distractor objects, so the multi version is forced to learn instruction-grounded visual recognition, while single trained on the generic "place the object in the basket" cannot ground.

Usage

python

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

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

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

SmolVLA RoboTwin place_object_basket (50 ep, MULTI instruction)

Task & Training

Evaluation

Usage

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

SmolVLA RoboTwin `place_object_basket` (50 ep, MULTI instruction)