pi0.5 speed ablation (speeda) — OpenArm + RH56F1, real teleop + synthetic, step 20k
LeRobot-native pi05 (v0.6.1) fine-tuned on real teleop v4 (261 ep, 20 Hz) + 324 synthetic RH56F1 episodes from the EE-speed ablation: A_norm — skeleton flow time-warped so the EE speed matches teleop (keyframe images unchanged).
8 synthetic tasks (airfryer_2, ball, bottle, bottle_over_shelf, bottle_pour_2, box, coffee_pot, doll); 4 real + 8 synthetic = 12 cells, shares ∝ sqrt(frames).
Synthetic actions: depth IDM + HaWoR wrist IK on the speed-warped videos.
Comparison points (same recipe): RyanL22/pi05-openarm-rh56f1-ablation-noskel-realsynth-20k, ...-firstkf-realsynth-20k, and the full pipeline RyanL22/pi05-anyh2r-rh56f1-0916-wristik-30k.
Checkpoint: step 20,000 (final), NAVER MLXP 4x H200, batch 16 x 4 = 64.
| value | |
|---|---|
| vision encoder (SigLIP) | fine-tuned (not frozen) |
| image augmentation | photometric + affine, one draw replayed across the stereo pair |
| mirror augmentation | off |
| optimizer | AdamW, peak lr 2.5e-5, cosine decay, warmup 1000 |
| chunk | 50 actions @ 20 fps (2.5 s) |
| normalization | quantiles; constant dims (head) widened to mean ± 0.1 rad |
Inputs: base_0_rgb <- left ZED 288x512, left_wrist_0_rgb <- right ZED; state/action 28 dims neck(2) | left_arm(7) | right_arm(7) | left_hand(6) | right_hand(6).
Rollout: feed the state neck as 0.889 / 0.001.
from lerobot.policies.pi05.modeling_pi05 import PI05Policy
policy = PI05Policy.from_pretrained("RyanL22/pi05-openarm-rh56f1-ablation-speeda-realsynth-20k")
- Downloads last month
- 22