CoRL2026-CSI/pi05_teleop_sort_block

π0.5 — SO-101 sort_blocks

Fine-tuned lerobot/pi05_base on 100 teleop episodes of the SO-101 sort_blocks task.

Model

• Architecture: π0.5 (PaliGemma-2B VLM + Gemma-300M action expert, flow matching, 10 inference steps)
• Cameras: base_0_rgb, left_wrist_0_rgb, right_wrist_0_rgb (224×224)
• State / Action dim: 32 (padded) / 6 (SO-101)
• Action chunk: 50
• dtype: bfloat16

Camera key rename (dataset → policy):

observation.images.top   → observation.images.base_0_rgb
observation.images.wrist → observation.images.left_wrist_0_rgb

right_wrist_0_rgb is an empty camera slot for this single-arm setup.

Action features (SO-101): shoulder_pan, shoulder_lift, elbow_flex, wrist_flex, wrist_roll, gripper (.pos). Normalization: ACTION/STATE = MEAN_STD, VISUAL = IDENTITY.

Data

CoRL2026-CSI/SO101-teleop_sort_blocks_100epi — 100 episodes, 94,568 frames, 30 fps, human teleop.

Training

Hardware	4 × GPU (DDP, 🤗 Accelerate)
Per-device batch	32
Gradient accumulation	2
Effective global batch	256
Steps	18,500 (~50 epochs)
Optimizer	AdamW, β=(0.9, 0.95), wd=0.01, grad clip 1.0
LR	cosine decay, peak 2.5e-5 → 2.5e-6, warmup 1000, decay 30000
Gradient checkpointing	on
Image aug	ColorJitter (brightness/contrast/saturation/hue), SharpnessJitter, RandomAffine — max_num=3, random order
Seed	1000

Training script: scripts/train_pi05_sort_blocks.sh.

Usage

from lerobot.policies.pi05.modeling_pi05 import PI05Policy

policy = PI05Policy.from_pretrained("CoRL2026-CSI/pi05_teleop_sort_block").to("cuda").eval()

lerobot-eval --policy.path=CoRL2026-CSI/pi05_teleop_sort_block --env.type=<env> --eval.n_episodes=20

Limitations

• Single task, single seed; no quantitative success rate reported here.
• Trained on a single-arm SO-101; the right-wrist camera slot is empty.
• 100 episodes only — sensitive to camera/lighting domain shift.

License

Apache 2.0 (inherits from lerobot/pi05_base).