SO_101_integration.md

SO-101 Integration Summary

This document summarizes the integration of the LeRobot SO-101 arm into the DISCOVERSE simulation environment.

Components

1. Environment (discoverse/robots_env/so101_base.py)

• Defines SO101Base and SO101Cfg.
• Handles the MuJoCo simulation interface for the SO-101 robot.
• Configures sensors (joint positions, velocities, end-effector pose) and actuators.
• Loads the robot model from mjcf/lerobot_so101/xml/so101_tabletop_manipulation_generated.xml.

2. Task Base (discoverse/task_base/so101_task_base.py)

• Defines SO101TaskBase, inheriting from SO101Base.
• Provides a foundation for robotic tasks, including methods for:
  • Resetting state (resetState)
  • Updating control signals (updateControl)
  • Checking action convergence (checkActionDone)
• Includes recoder_so101 for recording task data (actions and observations) to disk.

3. Inverse Kinematics (examples/mocap_ik/mocap_so101_ik.py)

• Implements SO101_IK using the Mink library.
• Solves for joint positions given a target end-effector position and orientation.
• Configured specifically for the 6-DOF SO-101 arm (5 arm joints + 1 gripper), targeting the gripperframe site relative to baseframe.

4. Example Task: Pick Milk (examples/tasks_so101/so101_pick_milk.py) THIS DOESN'T WORK PROPERLY YET, the frame transforms are off

• A complete example script utilizing SO101TaskBase and SO101_IK.
• Implements a state machine to:
  1. Hover over the milk object.
  2. Open the gripper.
  3. Descend to grasp position.
  4. Close the gripper.
  5. Lift the object.
• Demonstrates batch data generation and recording.

Limitations / Not Implemented

• Domain Randomization: The domain_randomization method in SO101TaskBase is currently a placeholder and needs to be implemented for robust policy training.
• Complex Interactions: Current success checks are simple (e.g., height check). More complex interaction logic might be needed for advanced tasks.