Strategic multi-agent interactions: game theory for robot learning and decision making

Conference on Robot Learning, 2022

Recording

Please find a recording of our workshop here: https://www.youtube.com/watch?v=8Ig2LYGvRuk

A separate recording of Drago Anguelov's talk, with improved video rendering, can be found here: https://www.youtube.com/watch?v=RpiN3LyMLB8

Schedule - December 15, 2022

Note: all times are in local NZ time. Video information can be found on the Pheedloop conference virtual platform .

0830: Opening remarks

0840: Invited talk (Jakob Foerster)

0915: Invited talk (Eric Mazumdar)

0950: Paper talks 1, 5, 8 + Q&A
- "Robust Forecasting for Robotic Control: A Game-Theoretic Approach," Shubhankar Agarwal (10 mins)
- "Collision Risk-based Congestion Game For Air Traffic Management," Sarah Li (10 mins)
- "Cautious Markov Games for Interaction Aware Robotics," Rohan Sinha (10 mins)

1035: Invited talk (Daniel Fried)

1110: Paper talks 3, 11, 14 + Q&A
- "Scalable Multi-Agent Reinforcement Learning through Intelligent Information Aggregation," Siddharth Nagar Nayak (10 mins)
- "Stateful active facilitator: Coordination and Environmental Heterogeneity in Cooperative Multi-Agent Reinforcement Learning," Dianbo Liu (10 mins)
- "Dominion: A New Frontier for AI Research," Danny Halawi (10 mins)

1155: Invited talk (Alyssa Pierson)

1230: Lunch break

1530: Paper talks 7, 9, 13 + Q&A
- "Iterative LQ Games for Occlusion Motion Planning," Kushagra Gupta (10 mins)
- "Autonomous Multi-Agent Racing using Constrained Potential Dynamic Games," Yixuan Jia (10 mins)
- "Nash Equilibria in Bayes Games for Coordinating with Imperfect Humans," Shray Bansal (10 mins)

1615: Invited talk (Drago Anguelov)

1650: Paper talks 4, 6 + Q&A
- "Game-Theoretical Perspectives on Active Equilibria: A Preferred Solution Concept over Nash Equilibria," Dong Ki Kim (10 mins)
- "Imperfect Information Games and Counterfactual Regret Minimization in Space Domain Awareness," Tyler Becker (10 mins)

1720: Closing remarks

Overview

To empower human activities and transform our lives, robots need to be deployed in environments shared by other intelligent decision makers. For instance, autonomous vehicles need to share roads with pedestrians, human-driven cars, and autonomous cars. Autonomous delivery drones need to navigate in airspace shared by other aircraft, and mobile robots in a warehouse must navigate factory floors shared by other robots as well as humans. To act appropriately in these complex scenarios, robots need to be capable of interacting strategically with other agents, which in turn requires us to develop interaction-aware, strategic learning and decision-making algorithms.

In this workshop, we study learning and optimal decision making in such strategic, multi-agent contexts. Although the game theoretical foundations of multi-agent strategic interactions are well-established, we aim to bring together expert researchers and practitioners with new and diverse experience and perspectives in the nascent field of game-theoretic decision making. In particular, we hope that workshop attendees will:

Gain a deeper technical understanding of the mathematical foundations of multi-agent games and a familiarity with modern algorithmic solution methods.
Engage in frank and illuminating discussions regarding the role of game-theoretic methods in robot learning
Be exposed to technical viewpoints from both within and outside the robotics community on challenging aspects of game-theoretic decision making, including:

High-dimensional, continuous state and action spaces
Partial observability
The subtle role of information structure
Algorithmic scalability

Areas of interest: Modeling multi-agent uncertainty, theory of mind, robust decision-making in dynamic environments, partial observability in multi-agent decision-making, etc.

Techniques include: game theory, optimization, optimal control, robust control, probability theory, bayesian inference, POMDPs, etc.