Agility Meets Stability: Versatile Humanoid Control with Heterogeneous Data скачать в хорошем качестве

Agility Meets Stability: Versatile Humanoid Control with Heterogeneous Data

Agility Meets Stability: Versatile Humanoid Control with Heterogeneous Data @ HKU | NVIDIA | Tsinghua We trained a unified whole-body control policy AMS for humanoid robots that enables zero-shot execution of diverse motions and behaviors with a single policy: 1. Dynamic motion tracking: Dancing 💃, running 🏃, etc. 2. Extreme balance control: Ip Man squat, single-leg balance, etc. 3. Real-time teleoperation: Low-latency real-time whole-body teleoperation Project page: https://opendrivelab.com/AMS/ arxiv: https://arxiv.org/abs/2511.17373 How to build a Behavior Foundation Model for humanoid robots? In this work, we tackle the data scale-up problem from a different angle: sampling directly from the robot's action space to generate reference motion data covering various possible balancing motions, rather than relying solely on manually collected human mocap data. This approach provides richer and more diverse reference motion data for policy training.