We present a system for generating physically plausible, whole-body humanoid motions on real hardware.

Leveraging the Unitree G1 platform, we deploy motions refined through optimization-based methods rather than reinforcement learning alone, enabling dynamic stability and realism.

See more details in LinkedIn post.

The robot learns to utilize its redundant degrees of freedom to determine actions for task execution.

Rather than relying solely on minimal necessary motions, it actively leverages redundancy to achieve more efficient and stable task performance.

The aerial manipulator hovers and switches to gimbaling mode for stable interaction.

The command center teleoperates the drone and manipulator to interact with an external object (light bulb).

To ensure compliant and safe interaction, the system integrates teleoperation with admittance control, and employs disturbance observers and external force estimation using low-cost servo motor sensors.