ucb_agentic_ai

Lecture 11: Autonomous Agents: Embodiment, Interaction, and Learning

Link to lecture recording on YouTube

Date: 2025-12-01

Speaker: Peter Stone

Speaker’s social profile: Website / University Profile / Google Scholar / GitHub / LinkedIn / X (Twitter)

Education:

• Ph.D. in Computer Science, 1998, Carnegie Mellon University
• B.S. in Mathematics, 1993, The University of Chicago

Work:

• Chair of Department of Computer Science, The University of Texas at Austin
• Chief Scientist, Sony AI

Notes

Agent: an intelligent system that interacts with the environment through sensors and effectors

• chatbot: input-output sequence, then wait for more input
• autonomous agent: continuous loop - sensation coming in from environment, actions taken that affect next sensations

Robot (physical / embodied agents): an intelligent system that interacts with the physical environment through sensors and effectors

Intelligent complete robot:

• perception (based on sensors’ reading from external world): sensing, modeling the world
• cognition: behaviors, action selection, planning, learning, teamwork etc.
• action (actuators take on external world): motion, navigation etc.

Speaker’s research question:

To what degree can autonomous intelligent agents learn in the presence of teammates and/or adversaries in real-time, dynamic domains?

AI systems have done better than the best humans at various tasks: chess, go, poker, StarCraft…, but soccer would be taking it to the next level in many ways

The speaker’s team published on the cover page of Nature¹, where they created reinforcement learning agents that were able to drive faster than the very best human drivers at the game Gran Turismo.
This was the first time that AI agents were better than humans at a physically realistic real-time control task, not turn-taking like poker, chess and go

[Incomplete, work in progress]

References

1. Nature: Volume 602 Issue 7896 (10 February 2022)

This site is open source. Improve this page.