Home
OmniXtreme Humanoid Robot Architecture Unveiled: Flow Matching Tackles High-Dynamic Movement Challenges
Unitree has officially released and open-sourced its humanoid robot motion control architecture, named OmniXtreme, accompanied by a technical paper authored by its founder, Wang Xingxing. This architecture comprehensively addresses the challenges of motion fidelity loss and physical instability when executing extensive motion libraries in highly dynamic environments, such as stage performances for humanoid robots.

The OmniXtreme architecture employs a two-stage training framework. The first stage, Scalable Flow-based Pretraining, distills complex, high-dynamic skills—including backflips, martial arts, and street dance—from various expert strategies into a unified model. It learns motion trajectories through generative flow matching, effectively overcoming the gradient conflicts typical in multi-task reinforcement learning. The second stage, Actuation-Aware Post-Training, incorporates residual reinforcement learning and detailed torque-speed envelope modeling. This allows the robot to autonomously adjust its actions based on real-time motor physical limits and regenerative power dynamics.
Experimental results indicate that this architecture achieved a 96.36% success rate for backflips on the Unitree G1 hardware platform, with end-to-end inference latency reduced to 10 milliseconds. The open-source release of OmniXtreme not only showcases the strong scalability of flow matching technology in embodied AI but also signifies a shift for humanoid robots from replicating isolated skills to possessing general, high-fidelity mobility. It establishes a crucial technical framework for the industry to develop robust control systems in complex real-world environments.
Related article
Apple Smart Glasses Could Debut at WWDC27, Highlighting Privacy Protection
Bloomberg’s Mark Gurman reports that Apple’s smart glasses, codenamed N50, are slated for a WWDC27 debut in June 2027, with a retail launch expected in autumn 2027. Originally targeted for late this year and early 2027, the device’s release has been
Inside Details Exposed About Next-Gen Gemini: Strained Computing Power, Internal Teams Disagreed on Development Priorities and Resource Allocation
Reports indicate that the launch of Google’s highly anticipated next-generation Gemini model has been pushed back. Internal disagreements over development priorities and resource allocation, combined with limited computing capacity and complex approv
OpenAI Dismisses Growth Slowdown Concerns, Says Multiple Business Units Accelerating
In response to external scrutiny regarding decelerating sales growth and missed internal benchmarks, AI leader OpenAI issued a confident statement on Tuesday, April 28. The company clarified that its consumer products and enterprise services are adva
Related Special Topic Recommendations
Comments (0)
0/500
Unitree has officially released and open-sourced its humanoid robot motion control architecture, named OmniXtreme, accompanied by a technical paper authored by its founder, Wang Xingxing. This architecture comprehensively addresses the challenges of motion fidelity loss and physical instability when executing extensive motion libraries in highly dynamic environments, such as stage performances for humanoid robots.

The OmniXtreme architecture employs a two-stage training framework. The first stage, Scalable Flow-based Pretraining, distills complex, high-dynamic skills—including backflips, martial arts, and street dance—from various expert strategies into a unified model. It learns motion trajectories through generative flow matching, effectively overcoming the gradient conflicts typical in multi-task reinforcement learning. The second stage, Actuation-Aware Post-Training, incorporates residual reinforcement learning and detailed torque-speed envelope modeling. This allows the robot to autonomously adjust its actions based on real-time motor physical limits and regenerative power dynamics.
Experimental results indicate that this architecture achieved a 96.36% success rate for backflips on the Unitree G1 hardware platform, with end-to-end inference latency reduced to 10 milliseconds. The open-source release of OmniXtreme not only showcases the strong scalability of flow matching technology in embodied AI but also signifies a shift for humanoid robots from replicating isolated skills to possessing general, high-fidelity mobility. It establishes a crucial technical framework for the industry to develop robust control systems in complex real-world environments.
Apple Smart Glasses Could Debut at WWDC27, Highlighting Privacy Protection
Bloomberg’s Mark Gurman reports that Apple’s smart glasses, codenamed N50, are slated for a WWDC27 debut in June 2027, with a retail launch expected in autumn 2027. Originally targeted for late this year and early 2027, the device’s release has been
Inside Details Exposed About Next-Gen Gemini: Strained Computing Power, Internal Teams Disagreed on Development Priorities and Resource Allocation
Reports indicate that the launch of Google’s highly anticipated next-generation Gemini model has been pushed back. Internal disagreements over development priorities and resource allocation, combined with limited computing capacity and complex approv
OpenAI Dismisses Growth Slowdown Concerns, Says Multiple Business Units Accelerating
In response to external scrutiny regarding decelerating sales growth and missed internal benchmarks, AI leader OpenAI issued a confident statement on Tuesday, April 28. The company clarified that its consumer products and enterprise services are adva











