RyanL22/pi05-openarm-rh56f1-ablation-noskel-realsynth-20k

pi0.5 ablation (noskel) — OpenArm + RH56F1, real teleop + ablation synthetic data, step 20k

LeRobot-native pi05 (v0.6.1) fine-tuned on real teleop v4 (261 ep, 20 Hz) + 324 synthetic RH56F1 episodes generated WITHOUT the skeleton-flow condition (no-skeleton-flow ablation); synthetic actions = raw depth-IDM labels (no wrist IK: the generated wrist does not follow the human wrist). Synthetic set: 8 tasks (airfryer_2 27, ball 43, bottle 60, bottle_over_shelf 47, bottle_pour_2 44, box 45, coffee_pot 25, doll 33), 81243 frames, 20 fps, 512x288. Merged: 4 real cells + 8 synthetic cells = 12 cells, shares ∝ sqrt(frames).

The comparison point is the full pipeline (skeleton-flow conditioned, all keyframes) trained the same way: RyanL22/pi05-anyh2r-rh56f1-0916-wristik-30k (and -15k).

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

Training settings

	value
synthetic action labels	raw depth IDM (no wrist IK)
vision encoder (SigLIP, 412.4M)	fine-tuned (not frozen)
image augmentation	photometric + affine, one draw replayed across the stereo pair
cell shares	∝ sqrt(frames), 12 cells
mirror augmentation	off
optimizer	AdamW, peak lr 2.5e-5, cosine decay, warmup 1000
precision	bfloat16, gradient checkpointing
chunk	50 actions @ 20 fps (2.5 s), n_obs_steps=1
normalization stats	quantiles; constant dims (head pitch/yaw) widened to mean ± 0.1 rad

Inputs

• observation.images.base_0_rgb <- left ZED view, 288x512; observation.images.left_wrist_0_rgb <- right ZED view
• observation.state / action — 28 dims: neck(2) | left_arm(7) | right_arm(7) | left_hand(6) | right_hand(6)
• Rollout: feed the policy state neck as 0.889 / 0.001 (training constant); the synthetic start pose is the same as the 0916 set.

Load

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