LGG100/ur3-test-32

pi05 LoRA — bimanual UR3, pick red cube onto white plate (action horizon 32)

openpi pi05 fine-tune of pi05_base on the cleaned LGG100/ur3-test data. Config and policy follow LGG100/ur3-test-lora; the one change is action_horizon 10 → 32.

Training

openpi config	pi05_ur3_test_lora_h32 (code/training/config.py)
init	gs://openpi-assets/checkpoints/pi05_base/params
LoRA	PaliGemma gemma_2b_lora + action expert gemma_300m_lora, no EMA
data	22 episodes / 3620 frames @ 15 Hz (one episode without a gripper release dropped)
steps	4000, batch 32, 1 × L40S
lr	cosine, warmup 200, peak 2.5e-5 → 2.5e-6
final loss	~0.010

Cleaning (code/scripts/clean_ur3_bimanual_lerobot.py): each episode starts at the first left-arm motion, still pauses of ≥3 frames before release are removed, and each episode ends 1 s (15 frames) after the left gripper releases.

Policy I/O

• Policy: aloha_policy with adapt_to_pi=False. Joints use delta actions; both grippers are absolute.
• state: 14 = left 6 joints + left gripper, right 6 joints + right gripper (rad; gripper 0 = open, 1 = closed)
• images: cam_high, cam_left_wrist, cam_right_wrist (RGB, 240×320 in training)
• prompt: pick up the red cube and place it on the white plate
• Output: 32 × 14 absolute joint targets (≈2.1 s at 15 Hz)

In this data the right arm and right gripper never move. In practice the policy is a left-arm policy.

Files

• params/, assets/ur3-test-clean/norm_stats.json, _CHECKPOINT_METADATA: openpi checkpoint (step 3999, no optimizer state)
• code/training/config.py: openpi config containing pi05_ur3_test_lora_h32
• code/policies/aloha_policy.py: input/output transforms (stock openpi)
• code/scripts/: hdf5 → LeRobot conversion and cleaning scripts

Serve

uv run scripts/serve_policy.py policy:checkpoint \
  --policy.config=pi05_ur3_test_lora_h32 --policy.dir=<this folder>