pi0.5 — anyh2r OpenArm + RH56F1, frozen SigLIP + stereo-consistent image aug (20k)
LeRobot-native pi05 fine-tuned on the 2026-09-13 anyh2r mixture (12 synthetic/IDM cells
from human video + 4 real teleop cells).
Checkpoint: step 20,000 of a 60,000-step run.
What is different from the previous run
previous (...-wristik-grasp-mirror-20k) |
this run | |
|---|---|---|
| vision encoder | fine-tuned | frozen (freeze_vision_encoder=true, 412.4M params) |
| image augmentation | none | photometric + affine, one draw replayed across the stereo pair |
| cell shares | equal per cell | ∝ sqrt(frames) (per-frame exposure spread 4.59x -> 2.14x) |
| mirror augmentation | p=0.5 | p=0.5 (unchanged) |
Why
The previous checkpoint chose its motion mode from the image domain: holding the robot state fixed and swapping only the image source (real teleop frame vs generated frame) changed the predicted action amplitude by 1.35-1.93x. That made rollouts on real camera images reproduce only 27-65% of the trained motion, while in-distribution chunk reproduction was ~1.00.
Freezing the vision tower removes that dependence.
| cross-domain probe (same state, image source swapped) | previous 20k | this 20k |
|---|---|---|
| ball | 1.48x | 1.16x |
| bottle | 1.87x | 0.91x |
| box | 1.93x | 1.00x |
| doll | 1.35x | 1.01x |
Absolute amplitude is preserved: in-distribution chunk reproduction stays at 0.85-1.09 of ground truth, and predicted amplitude on real rollout frames is unchanged or higher.
Inputs
observation.images.base_0_rgb<- left ZED view, 288x512observation.images.left_wrist_0_rgb<- right ZED view, 288x512observation.state— 28 dims:neck(2) | left_arm(7) | right_arm(7) | left_hand(6) | right_hand(6)- 20 fps, chunk 50 (2.5 s)
- Downloads last month
- 19