JingwuLuo

A π0.5 (openpi) policy fine-tuned for Unitree G123 + Dex1-1. The task is to pick up the leftmost object of the requested type in the target area. Only the left arm moves; the right arm stays at its initial pose. This is the openpi checkpoint from step 19999, the final step. It contains params/ (the weights) and assets/ (norm stats). The optimizer state (trainstate/) is not included. The model was trained on exactly two prompts. Pass them verbatim at inference time: The policy also accepts JPEG bytes for the images; they are decoded as BGR and converted to RGB. Images are resized to 224×224 inside the model. In the raw data, this is the concatenation of states/actions → leftarm.qpos(5) +…