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

pi0.5 baseline — EgoMimic-style masking (OpenArm + RH56F1), step 20k

LeRobot-native pi05 (v0.6.1) trained with the EgoMimic visual recipe: in both human and robot videos the arm is masked black (SAM3) and a red line is drawn along each arm; at rollout the robot's own images must get the same mask + line. Data = OpenArm teleop v4 (20 Hz, 261 ep in 4 cells) + anyh2r human video, 12 tasks / 466 episodes, whose 28-D actions come from dex-retargeting (Inspire config -> RH56F1)

• OpenArm arm IK (hand variant A). Dataset built by the EgoMimic baseline pipeline (egomimic_baseline_20260919).

Checkpoint: step 20,000 (final) of a 20,000-step run (Kakao, 4x H100, batch 16 x 4 = 64).

	value
vision encoder (SigLIP)	fine-tuned
image augmentation	photometric + affine, one draw replayed across the stereo pair
cells	16 (4 robot + 12 human), shares ∝ sqrt(frames)
mirror augmentation	off
optimizer	AdamW, peak lr 2.5e-5, cosine decay to 2.5e-6, warmup 1000
chunk	50 actions @ 20 fps (2.5 s)
normalization	quantiles; constant (neck) dims widened to mean ± 0.1 rad

Inputs: observation.images.base_0_rgb <- left ZED (288x512, masked + red line), observation.images.left_wrist_0_rgb <- right ZED, observation.state 28-D neck(2) | left_arm(7) | right_arm(7) | left_hand(6) | right_hand(6). Sibling baseline with a rendered-robot overlay instead: RyanL22/pi05-openarm-rh56f1-masquerade-baseline-20k.