摘要
High-fidelity motion tracking serves as the ultimate litmus test for generalizable, human-level motor skills. However, current policies often hit a 'generality barrier': as motion libraries scale in diversity, tracking fidelity inevitably collapses—especially for real-world deployment of high-dynamic motions. We identify this failure as the result of two compounding factors: the learning bottleneck in scaling multi-motion optimization and the physical executability constraints that arise in real-world actuation. To overcome these, we introduce OMNIXTREME, a scalable framework that decouples general motor skill learning from sim-to-real physical skill refinement. Our approach uses a flow-matching policy with high-capacity architectures to scale representation capacity without the interference-intensive multi-motion RL optimization, followed by an actuation-aware refinement phase that ensures robust performance on physical hardware. Extensive experiments demonstrate that OMNIXTREME maintains high-fidelity tracking across diverse, high-difficulty datasets. On real robots, the unified policy successfully executes multiple extreme motions, effectively breaking the long-standing fidelity–scalability trade-off in high-dynamic humanoid control.
结论
We presented OMNIXTREME, a two-stage framework for scalable high-fidelity humanoid motion tracking in high-dynamic regimes. By combining specialist-to-unified flow-based pretraining with actuation-aware residual reinforcement learning, OMNIXTREME mitigates both the learning bottleneck at scale and the physical executability bottleneck at sim-to-real deployment. Extensive simulation results show that OMNIXTREME preserves tracking fidelity substantially deeper into motion diversity than other baselines, and real-robot experiments demonstrate reliable execution of diverse extreme behaviors with a single unified policy, breaking the conventional fidelity–scalability trade-off.
For future research, jointly scaling data diversity and model capacity will be essential for enhancing the generalization of whole-body humanoid motor skills. As learning-based controllers are pushed toward more dynamic and hardware-constrained regimes, actuation-aware modeling becomes a critical component of the learning pipeline. By incorporating high-fidelity actuation characteristics—such as current, power, torque, and speed-dependent constraints—researchers can further bridge the sim-to-real gap, ensuring that learned behaviors translate seamlessly to physical humanoid robots.
一、论文核心定位与研究背景
1. 核心研究目标
论文旨在解决人形机器人领域长期存在的通用性壁垒：当运动库的多样性、动态难度提升时，现有控制策略的运动跟踪保真度会不可避免地崩溃，尤其在真实机器人部署的高动态场景中，形成了经典的保真度 - 可扩展性权衡困境。论文提出的 OmniXtreme 框架，通过两阶段训练范式，用单一统一策略实现了人形机器人多样化极端高动态动作的鲁棒控制，打破了这一长期存在的行业瓶颈。
