pi05-pointwam-assembly-chunk30-4gpu-b32-30k

pi0.5 — pointwam assembly (peg-in-hole)

RLWRLD OpenArm + RH56F1 bimanual platform finetune of pi0.5 (openpi PyTorch path). 30,000 steps on 4x H100, final loss 0.0003.

Training setup

Observation and action space

base_0_rgb        <- ego_left     320x180 -> 224x224
left_wrist_0_rgb  <- ego_right
right_wrist_0_rgb <- zeros, masked out (this platform has no third camera)
state / action    28-D absolute joint, padded to 32-D inside the model

What is in here

model.safetensors (7.5 GB), metadata.pt, and the normalization statistics under assets/. Optimizer state (13.5 GB) and intermediate checkpoints are not included.

An important caveat about the normalization statistics

openpi's compute_norm_stats accumulates in float32. For a dimension whose values sit far from zero but barely vary -- a joint angle around 1.0 with a spread of ~1e-3 -- E[x^2] - E[x]^2 cancels to exactly 0.0. openpi normalizes with (x - mean) / (std + 1e-6), so that zero becomes a division by 1e-6: |z| in the thousands, and on the action side it is squared into the loss, where it swamps everything else.

The assets/ in this checkpoint carry statistics with those dimensions recomputed in float64. Replacing them with output from a fresh compute_norm_stats run brings the defect back. Use what is bundled here; if you must recompute, check that no dimension has std == 0.

Evaluation notes

• batch size 32 is the total across ranks, not per GPU.
• Actions are absolute joint targets. There is no delta transform to undo.
• Do not request chunks longer than 30; the model never saw that range.

Data

36,857 frames / 64 episodes at 30 fps. Internal dataset, not public.