人形机器人的训练方式正经历重大转变。2026年9月1日，李飞飞创立的World Labs发布新一代世界模型Atlas1，该模型具有生成、重建与模拟场景的能力，可以复制和放大真机数据1。这一突破意味着机器人训练正从依赖真机远遥操作向仿真路线转变1。在早期的数据采集模式中，操作员穿戴VR头显远程驱动机器人进行数据收集1。
这一范式转变对中国已建立的超过120家数据采集中心构成重大冲击1。行业估算训练通用机器人需要千万甚至上亿小时的数据1，而目前真实交互数据仅有50万小时1。随着仿真技术的成熟，曾经依赖大规模真机操作积累数据的采集中心面临新的挑战，其未来发展方向亟待探索。
On September 1, 2026, World Labs, founded by Fei-Fei Li, unveiled Atlas, a next-generation world model designed to generate virtual environments and simulate real-world changes.1 This development marks a significant transition in humanoid robot training methodology, moving from reliance on physical teleoperation toward simulation-based approaches.1
The Atlas model possesses the capability to generate, reconstruct, and simulate scenarios while replicating and scaling data collected from real robots.1 According to industry estimates, training general-purpose robots requires tens of millions to hundreds of millions of hours of data, yet current real-world interaction data only reaches approximately 500,000 hours.1 The earlier paradigm of real-machine teleoperation involved operators wearing VR headsets to remotely control robots from a distance.1
This technological shift poses significant challenges for China's robotics sector, which has established over 120 data collection centers.1 The emergence of simulation-based training methods threatens the viability of these facilities, which were built under the previous teleoperation-dependent model.1
评论
还没有评论，欢迎留下第一条。