The Amazon Robotics team FAR has unveiled its inaugural research breakthrough in the field of humanoid robotics: OmniRetarget. This innovative robot operates independently of traditional sensing devices like cameras or radars. Remarkably, it can autonomously lift a chair weighing 9 jin (approximately 4.5 kg), ascend a 1-meter-high table, execute a somersault jump, and efficiently handle tasks such as moving boxes and navigating slopes.

The research behind OmniRetarget focuses on minimizing the Laplacian deformation energy between the source and target motions. This ensures that the robot's movements closely mimic human demonstrations while maintaining the integrity of local spatial structures and contact relationships. The algorithm achieves this by solving constrained non-convex optimization problems, thereby meeting a variety of stringent constraints.

Utilizing an interactive mesh-based approach, the system can seamlessly adapt to different robot forms and interaction scenarios, generating a wide range of trajectories. Furthermore, reinforcement learning techniques are employed to bridge dynamic differences, facilitating a zero-shot transfer from simulated environments to real-world hardware.

Experimental results showcase OmniRetarget's superiority in both kinematic quality and downstream policy performance, surpassing multiple open-source retargeting baselines.