smolvla_real_stack_red_blue_100

A SmolVLA policy for the SO-101 arm, fine-tuned from lerobot/smolvla_base. It was trained on the so101_real_stack_red_blue_100 dataset, which covers one task in the so101-libero-mini MuJoCo simulation.

• Task: Stack the red cube on the blue cube.
• Language instruction: "Put the red cube on top of the blue cube."
• Inputs: observation.state (6-D joint state) and three 256×256 RGB cameras (camera1, camera2, camera3)
• Output: 6-D action, predicted in chunks of 50
• Training: 100,000 steps, batch size 32, learning rate 0.0001, seed 1000

The weights at the root of this repo come from the final checkpoint (step 100,000).

Evaluation

The policy was run for 50 episodes per checkpoint in simulation. The 95% interval is a Wilson score interval.

Checkpoint step	Successes	Success rate	Wilson 95% CI
20,000	30/50	60%	[46.2%, 72.4%]
40,000	39/50	78%	[64.8%, 87.2%]
60,000	40/50	80%	[67.0%, 88.8%]
80,000	43/50	86%	[73.8%, 93.0%]
100,000	39/50	78%	[64.8%, 87.2%]

Each checkpoint's full results (results.json, episodes.csv and rollout videos) are in eval/.

Usage

from lerobot.policies.smolvla.modeling_smolvla import SmolVLAPolicy

policy = SmolVLAPolicy.from_pretrained("tsangb34/smolvla_real_stack_red_blue_100")