Kaminoikari/smolvla-so101-pickplace-ft

SmolVLA · SO-101 Pick-and-Place (fine-tuned)

Fine-tune of lerobot/smolvla_base (450M VLA) on the real-robot SO-101 dataset lerobot/svla_so101_pickplace (50 teleoperated episodes, 2 cameras up/side, 6-DoF state). lerobot/svla_so101_pickplace lerobot/svla_so101_pickplace lerobot/svla_so101_pickplace (50 teleoperated episodes, 2 cameras up/side, 6-DoF state).

Training

Base	lerobot/smolvla_base
Dataset	lerobot/svla_so101_pickplace (50 ep / 11,939 frames)
Steps	2,000 (batch size 8)
GPU	single T4 (~78 min)
Camera mapping	up to camera1, side to camera2 via --rename_map
Loss	0.410 to 0.141 (monotonic)

Intended use & honest limitations

This is a pipeline-validation / learning run, not a production policy.

• Demonstrates the full real-robot imitation-learning loop: load a real teleoperation dataset, fine-tune a pretrained VLA, converge, ship a checkpoint.
• Only 2,000 steps (~1.3 epochs). The SmolVLA paper uses ~20k steps; expect this checkpoint to under-perform a fully trained one.
• No closed-loop success rate. Evaluation on the physical SO-101 arm (lerobot-record) was not run (no hardware). Reported signal is training-loss convergence only, which proves the model is learning, not real-world task success.

Load

from lerobot.policies.smolvla.modeling_smolvla import SmolVLAPolicy
policy = SmolVLAPolicy.from_pretrained("Kaminoikari/smolvla-so101-pickplace-ft")