Modelpublic

RyanL22/pi05-openarm-rh56f1-teleopv4-baseline-20k

sourceHugging Faceapache-2.0updated 8d agoView on Hugging Face

0likes21downloads

Model Card

pi0.5 baseline — OpenArm + RH56F1, real teleop only (teleop v4, 20 Hz), step 20k

LeRobot-native pi05 (v0.6.1) fine-tuned only on real teleoperation data: OpenArm teleop v4, the 261 episodes recorded at 20 Hz (ball 70 / bottle 48 / box 88 / doll 55, 87,905 frames). No synthetic / human-video data. This is the teleop-only baseline for the anyh2r experiments.

Checkpoint: step 20,000 (final) of a 20,000-step run (lab-gpu26, 2x H100, batch 32 x 2 = 64).

Training settings

	value
vision encoder (SigLIP, 412.4M)	fine-tuned (not frozen)
image augmentation	photometric + affine, one draw replayed across the stereo pair
cell shares	∝ sqrt(frames), 4 cells
mirror augmentation	off
batch	32 per GPU x 2 GPUs = 64 (same global batch as the 16 x 4 runs)
optimizer	AdamW, peak lr 2.5e-5, cosine decay to 2.5e-6, warmup 1000
precision	bfloat16, gradient checkpointing
chunk	50 actions @ 20 fps (2.5 s), `n_obs_steps=1`
normalization stats	constant dims (head pitch/yaw) widened to mean ± 0.1 rad

Unlike the ...-0916-* checkpoints, the neck quantiles are widened here, so a sub-degree encoder offset on the (fixed) head no longer normalizes to a huge value. Overwriting the state neck with 0.889 / 0.001 is still harmless.

Inputs

—observation.images.base_0_rgb <- left ZED view, 288x512
—observation.images.left_wrist_0_rgb <- right ZED view, 288x512
—observation.state — 28 dims: neck(2) | left_arm(7) | right_arm(7) | left_hand(6) | right_hand(6)
—action — same 28-dim layout, absolute joint targets

Tasks: pick up the {ball | bottle | small box | doll} from table and put it in the white box.

Load

python

from lerobot.policies.pi05.modeling_pi05 import PI05Policy
policy = PI05Policy.from_pretrained("RyanL22/pi05-openarm-rh56f1-teleopv4-baseline-20k")