Modeling and Planning for Human-Robot Mutual Adaptation

Abstract:

Human-Robot Interaction (HRI) has been popular in recent years, since robots are coming into people's daily life, e.g., autonomous driving, homecare robots, e.t.c. In order to be a helper rather than a hindrance, the robot has to reason over human's behaviors, and plan its own actions accordingly.

In this thesis, we develop a mathematical model for human-robot mutual adaptation, and solve it in a computationally tractable manner. Our model consists of two parts: (1) a human behavioral model that predicts human actions. In particular, we identify human trust and human intention as two important latent variables for pre-dicting human behaviors; (2) a probabilistic planning algorithm for robot sequential decision making under uncertainty. While the partially observable Markov decision process (POMDP) provides a principled solution for planning under uncertainty, it is hard to scale to large problems. We introduce POMDP-lite, a factored model that restricts partial observability to a state variable that is constant or changes determin-istically, for efficient robot planning under uncertainty.

The proposed model has been successfully applied to a human-robot collaborative task, where we show that by explicitly modeling human trust, the robot is able to actively infer and influence human actions, which in turn improves team performance. To complete the thesis, we are working on the task of autonomous driving at unsignalized intersections, where we investigate the human driver's adaptation model by explicitly modeling human intentions.