arrow-hf/smolvla-robotwin-place-empty-cup-50ep

SmolVLA RoboTwin place_empty_cup (50 ep, single instruction)

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

This is the 50-ep counterpart to arrow-hf/smolvla-robotwin-place-empty-cup-300ep (300-ep, 90% success). Trained with the same hyperparameters as the other 50-ep tasks in this benchmark for fair comparison.

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

Training

Config	Value
Base checkpoint	lerobot/smolvla_robotwin
Training data	50 RoboTwin demonstrations (subset of 300ep_A), single instruction
Batch size	32
Steps	6000
Optimizer	AdamW, lr=1e-4
Scheduler	Cosine, warmup=300, decay=6000
Chunk size	50
Final train loss	0.005

Evaluation

RoboTwin 2.0 sim (demo_clean), 10 episodes, max_steps=400, action_chunk_exec=50, instruction "place the empty cup".

Success rate: 7/10 (70%)

Single vs Multi Comparison

Variant	Data	Success rate
50ep single (this model)	50	70%
50ep multi	50	80% ⭐
300ep single	300	90%

Interesting finding: At 50 episodes, the multi-instruction variant actually outperforms single (+10pp). This is one of two tasks in our 8-task benchmark where multi training improves over single (the other is place_object_basket).

Usage

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

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