Japanese Firm to Train Robot with AI to Reproduce Runners' Movements 日本公司利用AI训练机器人复现跑步者动作

TL;DR

• GMO Internet Group is launching a trial project to train humanoid robots to replicate the running movements of elite Japanese athletes using AI.
• The initiative involves quantifying motion data from professional runners and inputting it into an AI platform installed on Unitree’s G1 humanoid robot.
• This marks the world’s first attempt to autonomously replicate human running based on sensor data, with goals including winning events at the World Humanoid Robot Games.
• Current demonstrations show limitations, as the robot struggles with lane discipline and stability, falling during trials in Kawasaki.
• GMO aims to expand humanoid robot applications beyond sports to logistics and disaster rescue within a few years.

Why It Matters

This project highlights the convergence of biomechanics, high-performance sports science, and embodied AI, pushing the boundaries of how robots learn complex motor skills through imitation learning. For researchers, it demonstrates the practical application of transferring human kinematic data to robotic control systems, a critical step toward more agile and naturalistic robot movement. Industry observers should note the strategic use of competitive robotics events as a benchmark for technological advancement and public engagement.

Technical Details

• Data Acquisition: Motion capture data is collected from elite athletes, including Yuya Yoshida (marathoner) and Asahi Kuroda (ekiden runner), using sensors to quantify running mechanics.
• AI Platform: The collected kinematic data is processed by GMO’s proprietary AI platform, which learns to reproduce these specific movement patterns.
• Hardware: The AI is deployed on the Unitree G1 humanoid robot, which stands 1.3 meters tall, weighs 35 kilograms, and has a current top speed of 3 meters per second (approx. 100m in 30 seconds).
• Performance Status: Initial tests revealed significant challenges, with the robot failing to stay in lanes and experiencing falls, indicating that autonomous replication of human running is still in early developmental stages.
• Target Application: The immediate technical goal is to optimize gait and balance for competition in the World Humanoid Robot Games in China, while long-term goals include warehouse delivery and disaster response.

Industry Insight

The focus on replicating human athletic performance suggests a broader industry trend toward using high-dynamic tasks as stress tests for robotic stability and control algorithms. Success in this area could accelerate the deployment of humanoid robots in unstructured environments where agility and adaptability are paramount. Companies should monitor developments in imitation learning from biological sources, as this approach may offer faster convergence for motor skill acquisition compared to traditional reinforcement learning methods.

TL;DR

• GMO Internet Group启动试验项目，利用AI平台收集日本顶尖田径运动员的运动数据，训练人形机器人自主复现跑步动作。
• 该项目搭载于中国宇树科技（Unitree）开发的G1人形机器人，旨在通过传感器量化运动员姿态并输入AI模型进行模仿学习。
• 尽管演示中机器人出现偏离跑道和摔倒等情况，GMO仍计划参加8月在中国举行的世界人形机器人运动会田径赛事。
• GMO拥有10台人形机器人，短期目标是竞技获奖，长期愿景是将其应用于仓库物流、工厂作业及灾害救援等实际场景。

为什么值得看

该案例展示了“数据驱动+具身智能”在复杂动态运动控制领域的最新探索，特别是将专业运动员的生物力学数据转化为机器人控制策略的尝试。对于关注人形机器人商业化落地和运动控制算法的行业从业者而言，这提供了从实验室仿真走向真实物理环境交互的重要参考路径。

技术解析

• 数据获取与量化：项目核心在于采集日本田径队精英选手（如代表日本参加世锦赛马拉松的吉田雄也和箱根驿传选手黑田朝日）的运动数据，通过传感器将跑步姿态、步频、力度等生物力学特征进行数字化量化。
• AI平台与模型训练：GMO开发专用的AI平台处理上述数据，训练机器人自主复制人类跑步动作。这涉及模仿学习（Imitation Learning）或强化学习技术，试图让机器人理解人类运动的内在逻辑而非简单轨迹跟随。
• 硬件载体规格：执行端采用宇树科技（Unitree）的G1人形机器人，身高1.3米，重35公斤。目前硬件基础性能为最高时速3米/秒（百米约30秒），但在实时平衡控制和路径跟踪上仍面临挑战，演示中出现了摔倒和偏离跑道现象。
• 应用场景规划：除了竞技展示，技术最终指向通用任务执行，包括仓储配送、工厂搬运及灾难救援，强调机器人在非结构化环境下的适应性和稳定性。

行业启示

• 人机协作的新范式：体育竞技数据成为训练机器人的优质数据集，表明高技能人类专家的经验可以通过数字化手段高效迁移至机器人领域，加速特定任务的学习曲线。
• 硬件迭代与算法瓶颈并存：尽管硬件（如G1）已具备基本运动能力，但复杂动态平衡和精准模仿仍是难点。这提示行业需继续加强感知-决策-执行闭环的鲁棒性研究，特别是在真实物理干扰下的稳定性。
• 商业化路径清晰化：从“参加机器人奥运会”这种高关注度事件切入，逐步过渡到物流、救援等B端/G端实用场景，是具身智能公司验证技术并建立市场认知的有效战略。