InternRobotics/OpenHomie

What it solves

HOMIE addresses the difficulty of precisely and efficiently controlling a humanoid robot's full-body movements for complex loco-manipulation tasks. It removes the dependency on expensive motion capture (MoCap) data and complex inverse kinematics (IK) calculations, which often slow down teleoperation and limit the robot's operational workspace.

How it works

The project combines a reinforcement learning (RL) training framework and a low-cost hardware cockpit:

• RL Framework (HomieRL): Uses an RL-based approach with an upper-body pose curriculum, height tracking rewards, and symmetry utilization to train robots to walk, squat, and maintain balance while the upper body moves dynamically.
• Hardware Cockpit (HomieHardware): An isomorphic exoskeleton system consisting of exoskeleton arms, 15-DoF motion-sensing gloves, and a pedal for locomotion commands. Because the hardware is isomorphic to the robot, joint positions can be mapped directly from the exoskeleton to the robot without needing IK.
• Deployment (HomieDeploy): Provides the software bridge between the hardware cockpit and the robot (specifically tested with the Unitree G1).

Who it’s for

Researchers and developers working on humanoid robotics, teleoperation, and reinforcement learning for loco-manipulation.

Highlights

• Isomorphic Control: Direct joint mapping from exoskeleton to robot for faster and more accurate teleoperation.
• MoCap-Free: Eliminates the need for motion priors from motion capture data.
• Low Cost: Hardware system costs approximately $0.5k, significantly cheaper than professional MoCap systems.
• Full-Body Agility: Enables robust walking and rapid squatting to various heights while maintaining balance during upper-body movement.