In groundbreaking robotics research, Nvidia and Carnegie Mellon University have collaborated to develop a training framework called ASAP (Aligning Simulation and Real Physics), enabling humanoid robots like the Yushu G1 to perform agile, human-like athletic movements. This advancement bridges the gap between simulated environments and real-world physical dynamics, a challenge long faced in robotics.
Human-Like Movements: From Simulation to Reality
The Yushu G1 robot, showcased in the research, can now replicate complex human actions, such as Kobe Bryant-style basketball shots, Ronaldo-like spins, and even intricate gymnastic routines. This is achieved through ASAP, a method that aligns motion-tracking policies pre-trained in simulation environments with the real world’s physical dynamics.
The process consists of four key stages:
- Motion Tracking Pre-Training and Real Trajectory Collection: Human video motion data is converted into robot-compatible actions, pre-trained in simulations, and used to generate real-world trajectory data.
- Differential Action Model Training: Real-world motion data is used to minimize discrepancies between simulated states and real-world states, training a “differential action model.”
- Policy Fine-Tuning: By integrating the differential model into the simulator, the pre-trained policies are fine-tuned to better align with real-world dynamics.
- Real-World Deployment: The refined policies are applied directly to the physical robot, eliminating the need for the differential model.
This framework has been tested across multiple transition scenarios—such as IsaacGym to IsaacSim, IsaacGym to Genesis, and IsaacGym to the real world—with the Yushu G1 robot demonstrating remarkable stability and coordination in movements like side jumps and “James Silencer” poses.
Open-Source Accessibility
Both the research paper and the code for ASAP are open source, allowing the global robotics community to access and build upon these innovations. This transparency is expected to accelerate the development of humanoid robotics across various industries.
A Collaborative Effort by Chinese Researchers
The research team comprises 18 contributors, with four co-first authors—He Tairan, Gao Jiawei, Wenli Xiao, and Yuanhang Zhang—most of whom are Chinese. Their diverse educational backgrounds include institutions like Shanghai Jiao Tong University, Tsinghua University, and Carnegie Mellon University, highlighting the global and interdisciplinary nature of this project.
The Future of Humanoid Robotics
While the Yushu G1’s movements are impressively human-like, the robot is still in the early stages of development. Jim Fan, one of the lead researchers, envisions a future where humanoid robots might participate in events like the “2030 Humanoid Olympics.” For now, the Yushu G1 robot garners attention not just for its athletic prowess but also for its lighter moments, such as performing choreographed dance routines.
The advancements presented by Nvidia and Carnegie Mellon University mark a significant step toward a future where robots can seamlessly perform human-like actions, opening doors to applications in sports, healthcare, manufacturing, and beyond.