OmniXtreme is an open-source control framework designed for high-dynamic humanoid control, specifically pushing humanoids to hyperhuman limits. The framework achieves unified policy control across diverse extreme behaviors, enabling humanoid robots to perform complex motions.

The framework perfectly balances generative Flow Matching for extreme motion planning with strict physical envelope clipping to prevent mid-air motor burnouts. It includes power-safety regularization that explicitly penalizes excessive negative joint power to prevent unsafe energy absorption and transient braking loads during high-dynamic motions. The system also incorporates motor characteristic modeling for realistic simulation.

OmniXtreme uses a two-stage training approach. A unified base policy is trained via DAgger-based Flow Matching to aggregate diverse motion priors from different motion tracking experts. The base policy is then frozen while a residual policy is optimized under stringent motor constraints, extensive domain randomization, and power-safety regularization to bridge the sim-to-real gap.

The framework enables robust and agile control in physical environments, allowing humanoids to perform extreme behaviors like back handsprings, flips, kicks, rolls, and complex dance movements. It supports various motion sequences including attack combos, handstand walks, push-ups, and roll-flip combinations.

The deployment pipeline is real-time and executed entirely onboard, making it suitable for physical humanoid robot applications. The framework targets researchers and developers working with humanoid robotics, particularly those using Unitree G1 robots.