OmniXtreme

OmniXtreme's deployment pipeline is designed for real-time execution entirely onboard the humanoid robot, eliminating reliance on external computation. The inference speed enables closed-loop control at rates required for high-dynamic maneuvers, such as consecutive back handsprings or butterfly kicks. A key innovation is the power-safety regularization, which explicitly penalizes excessive negative joint power to prevent destructive energy absorption during extreme motions like landings or rapid direction changes. Motor characteristic modeling, studied in the ablation experiments, further refines control by accounting for the real motor's nonlinearities. Together, these features allow OmniXtreme to push robots to hyperhuman performance without sacrificing hardware longevity.

The overall workflow of OmniXtreme follows a two-stage training paradigm: pre-training a unified base policy via imitation learning, followed by post-training with reinforcement learning-inspired optimization for sim-to-real transfer. During pre-training, the DAgger-based Flow Matching method aggregates motion priors from multiple experts, effectively building a generalist policy that knows how to perform hundreds of different extreme behaviors. Then, in post-training, the base policy is frozen and a residual policy is optimized under realistic motor constraints and domain randomization. This separation of concerns allows the system to first learn diverse motion patterns, then adapt them to a specific robot platform. Deployment is fully onboard, meaning all computations run on the robot's embedded computer, enabling real-time control without network latency.

Concrete use cases demonstrated in the behavior gallery include a humanoid robot performing a breakdancing sequence with rapid-fire back handsprings, transitioning through hand-to-headstands, and executing webster flips consecutively. Another scenario shows the robot alternating pistol squats, demonstrating dynamic balance and strength. In attack combos, the robot delivers spinkicks and butterfly kicks with precision. These behaviors are executed by a single unified policy without switching between controllers. The outcome is a robot that can fluidly move from crawling to flipping to handstand walking, showcasing unprecedented versatility. Such capabilities are crucial for applications in entertainment, rescue, and exploration where robots must adapt to unstructured environments. The supplementary demos also include a 'Knocked out' scenario, indicating the policy can handle unexpected disturbances and falls.

OmniXtreme is primarily targeted at robotics researchers, especially those working on humanoid control and reinforcement learning. The open-source code is available on GitHub, and the project is detailed in an arXiv paper. While the specific tech stack is not explicitly mentioned, the implementation likely uses standard deep learning frameworks such as PyTorch for policy learning. The project is a collaboration between research institutions and Unitree Robotics, suggesting compatibility with Unitree humanoid platforms. Pricing is free and open-source. As a research framework, it accelerates the development of agile humanoid robots by providing a ready-to-use control baseline. In summary, OmniXtreme breaks the generality barrier in high-dynamic humanoid control, enabling a wide range of extreme behaviors from a single policy, and sets a new standard for versatility and robustness in legged locomotion.