Pete Florence: Dense Visual Representations, NeRFs, and LLMs for Robotics
The Gradient: Perspectives on AI
00:00
Robotic Manipulation
Andrei: I think one of the really cool bits for me with this is kind of what the showed is often of visual motor learning. And in the end to end approach, right, you just do that from image to control without any sort of induced structure or, you know, whatever prior. So another thing too is we were doing some of these demos with imitation learning and reinforcement learning with 30 or 50 layer resnets doing the vision part and then training the visual motor policies like together with that. A lot of people would instead call it deep learning, but really only kind of use a three or four or five layer or six layer network. But we had these quite large, big,
Transcript
Play full episode
