People I (Mostly) Admire 154. Can Robots Get a Grip?
65 snips
Mar 29, 2025 
Ken Goldberg, a UC Berkeley professor, dives into the fascinating realm of robotics. He discusses how robots struggle with seemingly simple tasks like grasping objects due to unforeseen complexities. The conversation highlights the blend of artistry and science in robotics, showcasing projects that marry technology and creativity. Goldberg also critiques the hype surrounding AI and robotics, emphasizing the need for realistic expectations. Plus, he explores how embracing randomness in data modeling can lead to groundbreaking advancements in the field.
AI Snips
Chapters
Books
Transcript
Episode notes
3D Vision Challenge
- While high-resolution cameras provide detailed 2D images, obtaining accurate 3D spatial representation remains challenging for robots.
- This 3D depth information is crucial for tasks like grasping, where precise interaction is necessary, unlike autonomous driving that prioritizes collision avoidance.
AlphaZero vs. Robots
- AlphaZero's rapid chess mastery stems from its ability to process vast amounts of data in a perfect information game.
- Physical robot tasks, unlike chess, limit learning speed as real-world actions and feedback collection are time-consuming and complex.
Dex-Net
- Dex-Net, a dexterity network, uses a large dataset of 3D objects and grasps, adding noise to mimic real-world imperfections, to train robots.
- This method achieved over 90% success rates in picking up objects, significantly surpassing previous approaches.
