Abstract:
While mobile robots reliably perform service tasks by accurately localizing and safely navigating while avoiding obstacles, they do not respond in any other way to their surroundings. In this work, we introduce two methods that enable the robots to be more responsive to their environment, including humans and other robots. The first algorithm enables the robots to be more responsive by equipping them with models of multiple tasks (or goals) and a way to interrupt a specific task and switch to another task based on observations. We show that our algorithm learns when to switch between tasks and requires less sensory computations compared to the naive approach of learning a combined model for all the tasks. The second algorithm enables each robot to learn a policy to cooperate with other agents in order to accomplish its individual goal. In a multi-agent setting, the optimal policy of a single agent is largely dependent on the behavior of other agents. We investigate the problem of multi-agent reinforcement learning, focusing on decentralized learning in non-stationary domains for mobile robot navigation and propose a curriculum-based strategy for learning interactive policies. We evaluate our approach on both an autonomous driving lane-change domain and a robot navigation domain and show that our approach outperforms a state-of-the-art approach and results in interactive policies.

Committee:
Manuela Veloso (Advisor)
Reid Simmons
Kris Kitani
Philip Cooksey