pi05-rby1-assembletissue-chunk30-4gpu-b32-30k

pi0.5 — RB-Y1 assemble-tissue

RLWRLD RB-Y1 bimanual platform finetune of pi0.5 (openpi PyTorch path). 30,000 steps on 4x H100, 7h 57m, final loss 0.0007.

Training setup

Robot

RB-Y1 + Wuji hand (rby1m_wujihand2). 54-D absolute joint targets, grouped as:

right_arm_joints  [0:7]     right_hand_joints [7:27]
left_arm_joints  [27:34]    left_hand_joints [34:54]
action uses the same four groups as *_command_joints

No head joints, and the hands are 20-D each. State and action group identically here -- worth stating because the sibling OpenArm/RH56F1 checkpoints do not (there the action side splits head from arm while the state side folds them together). Do not carry that caveat across.

Cameras are a stereo ego pair, ego_left / ego_right, 320x180.

Data

35,613 frames / 50 episodes at 30 fps, single task "Put the tissue roll onto the holder." Internal dataset, not public.

A companion run of the same task on the OpenArm/RH56F1 platform (28-D) exists with the same chunk, batch and step count, so the two robots can be read against each other.

action_dim 54 — three layers were trained from scratch

pi0.5 defaults to action_dim=32, which sizes action_in_proj, action_out_proj and state_proj. 54 does not fit, so this run raised action_dim to 54. The consequence, measured rather than assumed -- the training log named exactly these and nothing else:

action_in_proj.weight    checkpoint (1024, 32) -> model (1024, 54)
action_out_proj.weight   checkpoint (32, 1024) -> model (54, 1024)
action_out_proj.bias     checkpoint (32,)      -> model (54,)

Those three started from random init; the PaliGemma backbone and the action expert body loaded from pi05_base normally. Final loss landed at 0.0007, the same order as the 28-D sibling runs (0.0002-0.0003), so the cost appears small -- but it is a real deviation from how the base model was pretrained and worth knowing before comparing.

Note that transforms.pad_to_dim only pads and never truncates, so a 54-D vector against action_dim=32 fails loudly on shape rather than being silently cut.

Camera mapping

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)

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.

The bundled statistics were checked for the std == 0 failure mode that openpi's float32 accumulator can produce (E[x^2] - E[x]^2 cancelling for dimensions far from zero with tiny variance). This dataset is clean -- minimum std is 2.2e-02 for state and 2.7e-02 for actions -- but if you regenerate them, check for zero-std dimensions before training.

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.