Karol Hausman 分享了他从童年对机器人的迷恋到创立 Physical Intelligence 的历程，这家公司正在为现实世界的机器人构建 AI 大脑。
Karol Hausman shares his journey from childhood fascination with robots to founding Physical Intelligence, a company building an AI brain for real-world robots.
要点 · TL;DR
通用机器人模型胜过专用模型，像大语言模型一样随数据扩展。 Generalist robot models outperform specialists, scaling with data like LLMs.
真实世界数据对操作任务至关重要，模拟无法替代。 Real-world data is essential for manipulation; simulation falls short.
强化学习因基础模型带来的更好探索而复兴。 Reinforcement learning returns with foundation models enabling better exploration.
核心观点 · Key points
通用模型在机器人领域优于专用模型，类似于语言模型。 Generalist models outperform specialists in robotics, similar to language models.
真实世界数据对操作任务至关重要；模拟难以应对世界复杂性。 Real-world data is crucial for manipulation tasks; simulation struggles with world complexity.
强化学习因基础模型带来的更好探索而复兴。 Reinforcement learning is making a comeback due to better exploration from foundation models.
扩展现有方案可能足够；研究可改善扩展斜率。 Scaling existing recipes may suffice; research can improve scaling slope.
机器人需要高可靠性；物理世界对错误不容忍。 Robots need high reliability; physical world is unforgiving of mistakes.
智能可以补偿不精确的硬件，实现灵巧性。 Intelligence can compensate for imprecise hardware, enabling dexterity.
反共识 · Contrarian takes
机器人进展令人失望；预编程机器缺乏智能。 Robotics progress was disappointing; pre-programmed machines lacked intelligence.
哲学大多是错误的历史；少数见解成立。 Philosophy is mostly history of mistakes; few insights hold true.
沉浸式学习（如语言）优于显式规则方法。 Learning by immersion (like language) beats explicit rule-based approaches.
模拟对操作任务不如对运动任务有用。 Simulation is less useful for manipulation than for locomotion.
价值函数能在人类察觉前预测机器人失败。 Value functions can predict robot failure before humans notice.
伟大常在不经计划中产生，如《为什么伟大不能被计划》所述。 Greatness often arises without planning, as argued in 'Why Greatness Cannot Be Planned'.
本期章节 · Chapters（共 30）
童年对机器人的痴迷Childhood Fascination with Robots
可乐罐实验：突破The Coke Can Experiment: A Breakthrough
机器人悖论与物理智能使命The Paradox of Robotics and Physical Intelligence's Mission
机器人越多，模型越好The More Robots Deployed, the Better Models Get
早期影响与成长Early influences and upbringing
早期兴趣与机器人之路Early interests and path to robotics
哲学研究与反思Philosophy studies and reflections
斯宾诺莎与实在结构Spinoza and the structure of reality
网球内心游戏与无意识学习The Inner Game of Tennis and unconscious learning
机器人视觉的转折点A turning point in robotics vision
寻找智能机器人Finding intelligent robots
转向深度学习Switching to deep learning
信念确立的时刻The moment of conviction
泰勒·斯威夫特演示The Taylor Swift demo
机器人结合大语言模型Combining robotics with LLMs
创业条件Conditions for Starting the Company
机器人通用模型Generalist Models for Robotics
运动与操作的仿真Simulation for Locomotion vs Manipulation
强化学习回归Reinforcement Learning Comeback
耐心与商业化压力Patience vs. Commercialization Pressure
优化学习速率Optimizing for Rate of Learning
最陡学习率论点Theses on Steepest Learning Rate
无需新研究的扩展Scaling Without New Research
评估用仿真Simulation for evaluation
进展速度惊人Surprising speed of progress
机器人的可靠性与自我意识Reliability and self-awareness in robotics
AI 自我评估优于人类专家AI's Superior Self-Assessment vs Human Experts
物理智能的本体感觉与补偿Proprioception and Compensation in Physical Intelligence
书籍推荐：伟大不可计划Book Recommendation: Why Greatness Cannot Be Planned