Sergey Levine 谈 Physical Intelligence 机器人基础模型
“The universe is really big, and almost none of it is inside computers.”
Aditya Agarwal 分享了 Sergey Levine 在 SPC 的对话：Levine 早年在改造的 Google 会议室里让一屋子机器人整日抓取物体，积累了大量经验却缺乏先验知识，这一瓶颈最终催生了 Physical Intelligence。对话涵盖 Pi Zero 机器人基础模型、从 Pi 0.5 到 Pi 0.7 的推理与泛化能力演进，以及机器人数据难题和世界模型。
正文
“The universe is really big, and almost none of it is inside computers.”
When a problem is hard, make it bigger. @svlevine put a room full of robots in a repurposed Google conference room and had them grasp objects all day.
It worked. Then it hit a wall: lots of experience, no prior knowledge.
That wall became @physical_int.
This conversation at @spc was awesome. I learned a lot!
(00:28) From Video Games to Robotics: Sergey Levine's Journey (04:24) Founding Physical Intelligence: Why the Timing and Team Mattered (06:30) Inside Pi Zero: Building a Foundation Model for Robots (09:37) From Pi 0.5 to Pi 0.7: Teaching Robots to Reason and Generalize (16:45) What Makes a Great Robotics Researcher? (21:10) Solving Robotics' Data Problem and Building Learning Feedback Loops (28:04) The Future of Robotics: General-Purpose AI vs. Specialized Robots (33:24) World Models, Data Strategy, and Building a Robotics Research Culture