Adaptive and Efficient Models for Intent Recognition

Abstract

Assistive robots should have the ability to understand the intent of humans, predict their behavior, and plan to provide anticipatory assistance in complex real-life environments. In this thesis, we present adaptive and efficient algorithms for recognizing human intent.

We develop adaptive models for human intent recognition in a simulated search and rescue scenario. Humans vary widely in their behavior style due to different preferences, initial beliefs, internal world models, and planning mechanisms. A generic (non-adaptive) prediction model, therefore, has limited utility in this setting. Our adaptive model can recognize a rescuer's behavior patterns online and make better predictions. We show that adaptive models trained on a wide variety of simulated planning-based agents can transfer to humans and outperform generic models trained on limited human data.

We also present an efficient inverse reinforcement learning algorithm, called f-IRL, which directly optimizes a parameterized reward function to match the demonstrator's state distribution. We show that f-IRL can efficiently learn the demonstrator's intent - it can learn to imitate control policies from just a single demonstration. In addition, we show that the learned reward can be used to transfer policies to different dynamics.