Creating Generalist Robot Models by Physical Intelligence

​In this talk, our guest, Danny Driess, a research scientist at Physical Intelligence (Pi), will discuss a path towards creating generalist robot models — a robot policy that can solve any task in any environment.

First, Danny will discuss PaLM-E, one of the first large embodied AI models that was trained on general vision-language and robotic data. PaLM-E showed how knowledge from general tasks can be transferred into the embodied domain. PaLM-E focused on high-level reasoning.

Then he will explain how these findings can be transferred into training generalist low-level policies that output robot actions directly. This includes building RT-2, Pi0, and Pi0-FAST. In particular, with Pi0 the team showed one of the first generalist policies that can solve very long-horizon, dexterous tasks such as unloading a dryer and subsequently folding pieces of laundry, fully autonomously.

​Bio: Danny Driess is a research scientist at Physical Intelligence (Pi). Prior to that, he was a senior research scientist at Google DeepMind, working in the intersection between robotics and Gemini.

Join the BuzzRobot community Slack to stay in touch