Pi05-SO101-30

PI0.5 fine-tuned on the SO-101 black-cube-to-green-square dataset.

Training

• Base model: lerobot/pi05_base
• Dataset: so101_black_cube_green_square_30
• Local dataset root used for training: datasets/so101_black_cube_green_square_30
• Objective: pick up the black cube and place it in the green outlined square.
• Steps: 20,000
• Batch size: 1
• Precision: bfloat16
• Runtime: 12,368 seconds, about 3h 26m 08s
• GPU: NVIDIA A10G
• Git commit: 3bd9fb9d9114f49b6d173297d6ccbe95dba84c9e

Training used train_expert_only=true, freeze_vision_encoder=true, and gradient_checkpointing=true.

The SO-101 dataset camera keys were mapped for PI0.5 as:

• observation.images.top -> observation.images.base_0_rgb
• observation.images.wrist -> observation.images.left_wrist_0_rgb

Loading

In a compatible checkout of this repo:

uv run lerobot-eval --policy.path=coltonhabr/Pi05-SO101-30

The exported policy config points back to lerobot/pi05_base as its base model and includes the trained checkpoint weights and processor files.