VLA (3) pi0.5#

This document provides information about the pi0.5 VLA (Vision-Language-Action) implementation for XLeRobot with dual-arm support.

Overview#

The pi0.5 implementation includes optimizations for data collection scripts and dual-arm teleoperation, data collection, and inference workflows. Some changes are available in local branches for reference.

Implementation Details#

XLeRobot Fork#

  • Repository: xuweiwu/XLeRobot

  • Branch: pi05 dual arm

  • Features: Teleoperation, data collection, and inference for dual-arm setup

OpenPI Fine-tuning#

  • Repository: xuweiwu/openpi

  • Branch: biso101 training_support

  • Features: Training support for bimanual SO-101 configuration

Dataset#

The collected dataset for pi0.5 training is available on Hugging Face:

You can visualize the dataset using the LeRobot dataset visualization tool:

lerobot-visualize-dataset xuweiwu/bimanual-toy-box-cleanup

Alternatively, you can view it directly on Hugging Face using the lerobot/visualize_dataset space.

Status#

A minor PR has been created to optimize the existing data collection script workflow. Additional changes that may be more difficult to merge are currently available in local branches for reference.