Unitree officially released and open-sourced the humanoid robot motion control architecture named OmniXtreme, and simultaneously published a technical paper signed by its founder, Wang Xingxing. This architecture systematically solves the issues of fidelity degradation and physical landing when tracking a large action library in high-dynamic scenarios such as the Spring Festival Gala for humanoid robots.

QQ20260305-112924.png

OmniXtreme adopts a two-stage training framework: In the first stage, Scalable Flow-based Pretraining distills high-dynamic skills (such as backflips, martial arts, and street dance) from different expert strategies into a unified model. It learns velocity field paths through generative modeling, effectively avoiding gradient interference in multi-task scenarios with traditional reinforcement learning. In the second stage, Actuation-Aware Post-Training is introduced, using residual reinforcement learning and realistic torque-speed envelope modeling to allow the robot to self-correct based on motor physical limits and regenerative power.

Experimental data shows that this architecture achieved a 96.36% success rate for backflips on Unitree G1 hardware, with end-to-end inference latency compressed to 10 milliseconds. The open-source release of OmniXtreme not only demonstrates the excellent scalability of flow matching technology in embodied intelligence, but also marks a transition of humanoid robots from single-skill replication to general, high-fidelity mobility capabilities, providing a key technical paradigm for the industry to explore robust control in complex physical environments.