real_robot/source/examples/Xtrainer2/README.md

Xtrainer2 data adapter

This directory contains optional StarVLA configurations and conversion helpers for Xtrainer2 LeRobot data. It does not distribute robot data, checkpoints, or on-robot control software. The ReTime command in the repository root is the recommended entrypoint for proposal generation and training.

The supplied configurations cover pack-phone with absolute end-effector actions and pick-bottle with relative end-effector actions. For bimanual data, the source action contains 16 values: position, quaternion, and gripper for each arm. Rotation conversion expands the model action to 20 values. pytorch3d is required for this conversion; install a build matching your Python, PyTorch, and CUDA environment. The additional Python dependencies are listed in requirements-xtrainer2.txt.

Set dataset, model, statistics, training budget, and process count for your own environment. Keep the trainer gradient accumulation equal to the Accelerate configuration. Use a full checkpoint directory with --start resume when optimizer and scheduler continuation is required. A model file with the default --start weights initializes a new optimizer and scheduler.

Policy serving uses the included deployment/model_server implementation. A separate client must supply observations and consume actions using the same schema as the selected data configuration. Client source and deployment instructions are outside this artifact. See ROBOT.md for the data and interface contract.