Negar Mehr | Interactive Autonomy: Learning and Control for Multi-Agent Interactions

Series overviews and links can be found on our webpage: https://theairlab.org/tartanplannings...

Abstract:
To transform our lives, robots need to interact with other agents in complex shared environments. For example, autonomous cars need to interact with pedestrians, human-driven cars, and other autonomous cars. Autonomous delivery drones need to navigate in the aerial space shared by other drones, or mobile robots in a warehouse must navigate in the factory space shared by robots. The interactive nature of such application domains requires us to develop a systematic methodology for enabling efficient interactions of robotic systems across various applications. The goal of my research is to develop algorithms and mathematical models that enable safe and intelligent interactions in such multi-agent domains.

In this talk, I will first focus on game-theoretic planning and control for robots. To reach intelligent robotic interactions, robots must account for the dependence of agents' decisions upon one another. I will discuss how game-theoretic planning and control enables robots to be cognizant of their influence on other agents. I will present our recent results on leveraging the structure that is inherent in interactions to develop efficient motion planning algorithms which are suitable for real-time operation on robot hardware. In the second part of the talk, I will focus on how robots can learn and infer the intentions of their surrounding agents to account for agents' preferences and objectives. Currently, robots can infer the objectives of isolated agents within the formalism of inverse reinforcement learning; however, in multi-agent domains, agents are not isolated, and the decisions of all agents are mutually coupled. I will discuss a mathematical theory and numerical algorithms for inferring these interrelated preferences from observations of agents’ interactions.