How OmniXtreme Breaks the High‑Dynamic Control Barrier for Humanoid Robots

Background and Motivation

Achieving human‑level motion skills in general‑purpose humanoid robots has long been a research goal, but existing multi‑action control strategies often fail when faced with large, diverse motion libraries and the physical realities of hardware, leading to tracking collapse and unsafe deployments.

Key Challenges

Simulation‑to‑real transfer bottlenecks caused by limited parameterized policies that cannot handle heterogeneous actions and contact patterns.

Gradient interference in reinforcement‑learning (RL) when training a single policy for many actions, resulting in overly conservative behavior.

Physical hardware limits such as torque‑speed envelopes, regenerative power spikes, and sudden braking loads that are ignored in simplified simulators.

OmniXtreme Framework

OmniXtreme addresses these obstacles with a dual‑stage pipeline:

Scalable Flow‑Based Pretraining : A generative flow‑matching model is trained on expert policies derived from PPO for each motion in large datasets (LAFAN1, AMASS) that have been retargeted to the Unitree G1 robot. The flow model learns a velocity field guiding states toward expert actions rather than memorizing joint trajectories.

Actuation‑Aware Post‑Training : A lightweight residual MLP refines the pretrained policy on real hardware. It receives the same proprioceptive observations plus the previous refined action, and is trained with PPO rewards that penalize unsafe torque‑speed usage and excessive negative mechanical power.

The pretraining stage builds a unified behavior foundation, while the post‑training stage adapts it to the robot’s actual dynamics.

Flow‑Matching Details

The observation space includes joint positions, velocities, base angular velocity, and a 6‑D torso orientation error. Historical proprioceptive data are also provided. During training, the model samples time steps from a beta distribution to focus on critical trajectory regions, and integrates the learned velocity field with forward Euler steps. Randomized noise and domain randomization are injected to improve robustness.

Physical Deployment Strategy

In the post‑training phase, the pretrained flow policy is frozen. The residual MLP adds small corrective actions that are summed with the base output. Additional observations include the previous refined action. The training incorporates aggressive domain randomization (increased pose noise, force disturbances, terrain variations) and a power‑safety regularizer that caps negative mechanical power at each joint, protecting especially the knee during high‑impact landings.

Realistic torque‑speed envelopes and nonlinear friction are modeled in the simulator, ensuring that commanded torques never exceed what the hardware can deliver.

Experimental Evaluation

Simulation experiments show that OmniXtreme maintains low kinematic error and high success rates across a 60‑motion extreme dataset (XtremeMotion), outperforming both expert‑to‑MLP distillation and end‑to‑end RL baselines, whose performance degrades sharply as the motion set expands.

Real‑world tests on the Unitree G1 involved 157 trials covering 24 distinct high‑dynamic actions. The system achieved a 96.36% success rate on backflips and comparable performance on dance and martial‑arts motions, demonstrating that the simulation gains translate without loss.

Ablation studies reveal that each component—flow‑matching pretraining, residual post‑training, aggressive domain randomization, and power‑safety regularization—contributes uniquely to handling specific failure modes such as torque overload, contact instability, and energy spikes.

Scaling experiments indicate that increasing model capacity directly improves tracking quality for the flow‑matching architecture, whereas traditional MLPs hit a capacity ceiling early.

Conclusion

OmniXtreme proves that a carefully decoupled two‑stage learning pipeline can break the long‑standing fidelity‑scalability trade‑off in humanoid robot control, delivering robust, high‑dynamic motion execution on real hardware while preserving safety and efficiency.