arrow-hf/smolvla-robotwin-place-bread-skillet-50ep

SmolVLA RoboTwin place_bread_skillet (50 ep, single instruction)

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

Task

Dual-arm pick-and-place: pick up the bread and place it inside the skillet.

• Robot: Agilex dual-arm, end-effector control (16D state, 16D action)
• Cameras: 3 RGB streams — dual_cam_global, cam_wrist_65, cam_wrist_75
• Control rate: 10 Hz
• Single fixed instruction: "place the bread in the skillet" (Strategy A, not random per-episode)

Training

Config	Value
Base checkpoint	lerobot/smolvla_robotwin
Training data	50 RoboTwin demonstrations (subset of place_bread_skillet_300ep), strategy A single instruction
Frames	8,298 (~165 frames/ep)
Batch size	32
Steps	6000 (~23 epochs)
Optimizer	AdamW, lr=1e-4
Scheduler	Cosine, warmup=300, decay=6000
Chunk size	50
Final train loss	0.008
Walltime	~2h 15min (A100)

Evaluation

Evaluated in RoboTwin 2.0 simulator (demo_clean config), 10 episodes, max_steps=400, action_chunk_exec=50.

Model	Data	Base	Success
SmolVLA (smolvla_base)	300 ep	smolvla_base	0/10 (0%)
SmolVLA (this model)	50 ep	smolvla_robotwin	6/10 (60%)
X-VLA (xvla-base)	300 ep	xvla-base	8/10 (80%)

Training with the smolvla_robotwin base checkpoint enables strong data efficiency: with 6× less data and 3× fewer training steps, this model jumps from 0% to 60% success rate on this task.

Successful episodes complete in 132–201 environment steps (13–20s); failed episodes time out at 400 steps.

Usage

from lerobot.policies.smolvla import SmolVLAPolicy

policy = SmolVLAPolicy.from_pretrained("arrow-hf/smolvla-robotwin-place-bread-skillet-50ep")

See LeRobot documentation for inference setup.

Citation

Built on SmolVLA and SmolVLA-RoboTwin, fine-tuned on data collected from RoboTwin 2.0.