Unsupervised Learning Ep 70: Karol Hausman and Danny Driess (Physical Intelligence) Unpack the Most Recent Breakthroughs & Path to Generalist Robots
57 snips
Jul 8, 2025 
In this insightful discussion, Karol Hausman, Co-founder and CEO of Physical Intelligence, and Danny Driess, Research Scientist, dive into the evolving realm of AI robotics. They share the transition from traditional coding to machine learning models capable of multitasking. Topics include robots folding laundry and the importance of effective data collection. They dissect the future potential of generalist robots and the challenges of integrating advanced models. Their conversation highlights the transformative impact of robotics on daily life and the importance of open-source collaboration.
AI Snips
Chapters
Transcript
Episode notes
From Hard-Coding to Learning
- Early robotics relied on hard-coded rules, making it impossible to solve complex tasks broadly.
- Learning-based methods, where robots learn from experience, emerged as a better paradigm over the past decade.
Generalist Models and Pretraining
- Large models trained on many tasks enable robots to generalize across tasks.
- Using pre-trained vision-language models reduces the robotics data needed by leveraging internet-scale knowledge.
The Taylor Swift Moment
- The "Taylor Swift" robot demo seemed simple but proved the power of fusing internet knowledge with robotic action.
- That moment convinced the team robotics models could harness pre-training for physical tasks.