VLA-JEPA-SO101-30

Fine-tuned VLA-JEPA policy for an SO-101 pick-and-place task:

Pick up the black cube and place it in the green outlined square.

Training

• Base model: lerobot/VLA-JEPA-Pretrain
• Dataset: local so101_black_cube_green_square_30
• Episodes: 30 teleoperated runs
• Steps: 5,000
• Batch size: 1
• Hardware: NVIDIA A10G
• Runtime: 32m48s
• Hub push during training: disabled

The training run used the SO-101 6D state/action space. The VLA-JEPA action/state adapter tensors were reinitialized to match that dimensionality.

Loading

From a compatible LeRobot checkout:

uv run lerobot-eval \
  --policy.path=coltonhabr/VLA-JEPA-SO101-30

For training or evaluation pipelines that use the original dataset camera keys, apply the same rename map used during fine-tuning:

rename_map:
  observation.images.top: observation.images.exterior_1_left
  observation.images.wrist: observation.images.exterior_2_left