Add model card

3aab4e6 verified 2 months ago

2.78 kB

license: apache-2.0
library_name: lerobot
pipeline_tag: robotics
tags:
  - lerobot
  - robotics
  - pi05
  - so101
  - dual-arm
  - imitation-learning
datasets:
  - CoRL2026-CSI/SO101-DualArm-teleop_fold_towel_100epi
base_model: lerobot/pi05_base

π0.5 — SO-101 Dual-Arm `fold_towel`

Fine-tuned lerobot/pi05_base on 100 teleop episodes of the dual-arm SO-101 fold_towel 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) / 12 (dual-arm SO-101: 6 left + 6 right)
Action chunk: 50
dtype: bfloat16

Camera key rename (dataset → policy):

observation.images.left_top    → observation.images.base_0_rgb
observation.images.left_wrist  → observation.images.left_wrist_0_rgb
observation.images.right_wrist → observation.images.right_wrist_0_rgb

Action features (12 DoF):

left_{shoulder_pan, shoulder_lift, elbow_flex, wrist_flex, wrist_roll, gripper}.pos
right_{shoulder_pan, shoulder_lift, elbow_flex, wrist_flex, wrist_roll, gripper}.pos

Normalization: ACTION/STATE = MEAN_STD, VISUAL = IDENTITY.

Data

CoRL2026-CSI/SO101-DualArm-teleop_fold_towel_100epi — 100 episodes, 65,151 frames, 30 fps, human teleop.

Training


Hardware	4 × GPU (DDP, 🤗 Accelerate)
Per-device batch	32
Gradient accumulation	2
Effective global batch	256
Steps	12,800 (~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_fold_towel.sh.

Usage

from lerobot.policies.pi05.modeling_pi05 import PI05Policy

policy = PI05Policy.from_pretrained("CoRL2026-CSI/pi05_teleop_fold_towel").to("cuda").eval()

lerobot-eval --policy.path=CoRL2026-CSI/pi05_teleop_fold_towel --env.type=<env> --eval.n_episodes=20

Limitations

Single task, single seed; no quantitative success rate reported here.
Trained on dual-arm SO-101; not directly transferable to single-arm setups.
100 episodes only — sensitive to camera/lighting domain shift.

License

Apache 2.0 (inherits from lerobot/pi05_base).

π0.5 — SO-101 Dual-Arm fold_towel