My research focuses on applying imitation learning techniques to difficult decision-making problems in robotics. In many domains, explicitly deriving an algorithm, policy, or utility mapping to solve the problem may be difficult. However, a human expert may be able to provide a good solution or quickly gain intuition about the form of the desired solution. Using techniques from Imitation Learning and Inverse Optimal Control, we can train these complex systems to approach human-level performance.
One instance of this problem is following natural language directions through unknown environments. This problem requires understanding the structure of language, mapping verbs and spatial relationships onto actions in the world, recognizing diverse landmarks located in the environment, and reasoning about the environment and landmarks that have not yet been detected. We have framed this problem as sequential decision making under uncertainty, and use expert demonstrations to learn a policy which predicts a sequence of actions that follow the directions.
In a related vein of research, I studied multi-robot task allocation for complex tasks and environments. Here the issue is learning a good utility function used by the task allocation mechanism, especially in the face of rich environment dynamics or an adversary team that is not fully modeled. We again used expert demonstrations to learn a policy which biases prices in a market-based task allocation mechanism, and applied this to a simulated disaster response scenario and an adversarial strategic game.
|The Robotics Institute is part of the School of Computer Science, Carnegie Mellon University.|
Contact Us | Update Instructions