宇树机器人能翻跟斗，跑马拉松却摔得东倒西歪，为何差别这么大？

Core Viewpoint - The discussion surrounding the Yushu H1 robot's ability to perform complex gymnastic movements while struggling in a half marathon highlights the "impossible triangle" of hardware, algorithms, and scenarios in the humanoid robotics field [2] Group 1: Hardware Design Challenges - The Yushu H1 features a joint motor torque density of 230Nm/kg and a real-time control system with millisecond response speed, allowing it to switch its center of gravity in 0.5 seconds, prioritizing instantaneous power output [3] - However, this design leads to a significant power consumption of 300W per joint during dynamic activities, which is much higher than during normal walking [3] - During a marathon at a speed of 6km/h, the robot requires over 500W of total heat dissipation, exceeding its passive cooling system's capacity, resulting in knee joint temperatures exceeding 80°C after one hour, causing a 23% decrease in torque precision [4] Group 2: Algorithmic Layers - The VSLAM (Visual Simultaneous Localization and Mapping) system developed by Yushu allows for environmental perception at 60 frames per second, with a gait generator trained through reinforcement learning, achieving a reaction delay of only 80ms when avoiding obstacles, close to human spinal reflex speeds [5] - For marathon running, an energy consumption model must be established, requiring the robot to keep energy use below 238Wh per kilometer to cover 21 kilometers with a total battery capacity of 5kWh [7] - If the path planning algorithm is not optimized, actual endurance may drop to 15 kilometers, and the robot's "5-minute battery swap" plan results in loss of motion memory, effectively resetting its learning model every 5 kilometers [7] Group 3: Differences in Participants - The Yushu H1 robot participating in the marathon was modified by a client and did not use the original factory algorithm, which is based on 100,000 hours of simulation training with a foot pressure sensor error compensation frequency of 200Hz [8] - In contrast, the third-party algorithm used had a sensor calibration frequency of only 50Hz, leading to a fourfold increase in error accumulation during continuous running, explaining why the robot could stand on one leg for over 30 minutes in the lab but struggled in the race [8] Group 4: Industry Insights - The trend towards performance specialization is evident, with Tian Gong Ultra reducing weight by 8kg and lowering energy consumption by 15% for inspection and logistics, while Yushu H1 retains a 12-degree-of-freedom dexterous hand for flexibility in rescue and service scenarios [9] - Technological breakthroughs are being pursued, such as Boston Dynamics' Atlas experimenting with liquid metal cooling systems to reduce joint operating temperatures by 40%, and Tesla's Optimus improving energy efficiency through silicon carbide inverters [9] - Over the next five years, humanoid robots are expected to achieve breakthroughs in both gymnastics and marathon capabilities as solid-state battery energy density surpasses 400Wh/kg [9] Conclusion - The evolution of technology is a necessary path, as seen in the transition from laboratory demonstrations to practical applications in humanoid robotics, emphasizing the importance of algorithm optimization for performance leaps [10]