SmolVLA SO-101 Tape Dispenser

This is a SmolVLA policy fine-tuned with behavior cloning on 20 teleoperated demonstrations for a real SO-101 tape-dispenser pick-and-place task.

Task prompt:

Put the blue tape dispenser in the red box.

Results

The fine-tuned policy completed 5/8 evaluation rollouts with small object position and rotation changes. Under the same robot, cameras, object, and prompt, the base SmolVLA checkpoint completed 0/8 rollouts.

This is a small real-robot fine-tuning experiment, not a robust benchmark.

Setup

• Robot: SO-101 leader/follower setup
• Cameras: 2 UVC cameras, wrist + side
• Base model: lerobot/smolvla_base
• Demonstrations: 20 teleoperated episodes
• Image resolution: 640x480

Limitations

The main failure mode was recovery after a missed grasp. Once the object was pushed into an unseen pose, the policy often failed to re-grasp it.

Repository

Code and rollout videos: https://github.com/ajingu/SmolVLA-SO101-Tape-Dispenser