中国团队新突破！HUSKY框架让人形机器人成了“滑板高手”！

Core Viewpoint - The article discusses the development of the HUSKY physical perception full-body control framework, which enables humanoid robots to effectively control skateboards, demonstrating advanced dynamic control capabilities in various environments [1][3]. Group 1: Technical Challenges - Controlling a humanoid robot on a skateboard involves complex dynamic control, as the skateboard is an underactuated wheeled platform subject to nonlinear dynamics [3][5]. - Traditional trajectory planning methods are inadequate due to the complex motion characteristics of skateboards, making real-time control challenging [5][8]. Group 2: HUSKY Framework Development - The HUSKY framework begins by analyzing the physical properties of the skateboard, establishing a foundational physical model that informs all subsequent control strategies [7][11]. - The framework incorporates adversarial motion prior (AMP) technology, allowing the robot to learn from human skateboarding data, significantly improving its performance [9][10]. Group 3: Control Strategies - The framework employs a physical-guided steering strategy, where the robot uses body tilt to control the skateboard's direction, enhancing precision and reducing effort [11][13]. - A trajectory-guided transition mechanism is designed to facilitate smooth transitions between pushing and steering phases, addressing a common stability issue in robotic skateboarding [16][19]. Group 4: Real-World Application - The HUSKY framework successfully bridges the gap between simulation and real-world application by accurately identifying skateboard physical parameters and employing domain randomization during training [20][22]. - In real-world tests, the robot demonstrated stable operation at a control frequency of 50Hz, effectively adapting to various surfaces and maintaining balance under external disturbances [23][24]. Group 5: Future Implications - The HUSKY framework not only enhances humanoid robot capabilities in skateboarding but also provides a reusable approach for solving complex control problems in dynamic environments [24]. - Future developments may include integrating visual state estimation for closed-loop control and adapting to more complex terrains, further improving the robot's adaptability [24][25].