A Bayesian Model of Interactive Robotic Training for Complex Tasks
Date: Tuesday, August 10, 2021
Time: 1:30 PM to 3:00 PM Note: all times are in the Eastern Time Zone
Location: Seminar Room D463 (Star)
Event Type: Thesis Defense
Host: Ankit Shah, MIT
Contact: Ankit Shah, email@example.com
firstname.lastname@example.org, email@example.com, firstname.lastname@example.org
TALK: Thesis Defense: Ankit Shah - A Bayesian Model of Interactive Robotic Training for Complex Tasks
Abstract: Domains such as high-mix manufacturing, domestic robotics, space exploration, etc., are key areas of interest for robotics. Yet, the difficulty of anticipating the role of robots in these domains is a crucial hurdle for the adoption of robots. Developing robots that can be re-programmed easily during deployment by domain experts without requiring extensive programming knowledge is a key research thrust of this dissertation.
I present an online, multi-modal Bayesian framework for teaching a robot learner to perform the teacher's intended task from demonstrations and acceptability assessments of the robot's task execution provided by the teacher. I utilize a well-defined fragment of linear temporal logic (LTL) as the task specification language to enable robots to learn temporally extended tasks. In developing this framework, I address three key research questions.
I begin by presenting a novel approach to inferring formal temporal specifications from the observation of labeled task executions, called Bayesian specification inference. This approach can learn tasks belonging to an expressive but relevant fragment of LTL while modeling the ambiguity of demonstrations as a belief distribution over candidate LTL formulas. We demonstrate the utility of this approach in inferring task specifications for a representative multi-step manipulation task of setting a dinner table. We also utilize this model to learn an assessment model for multi-aircraft combat missions that shows a high degree of alignment with the assessments provided by a domain expert.
Next, I present planning with uncertain specifications (PUnS), a novel planning problem formulation that admits a belief distribution over the true specification. I propose four evaluation criteria that capture the semantics of satisfying a belief over logical formulas and demonstrate the existence of an equivalent Markov decision process (MDP) for every instance of a PUnS problem. We show that the robot policies produced through the PUnS formulation demonstrate flexibility by generating distinct valid task executions and result in a low error rate by simultaneously satisfying a maximal subset of the specifications in the belief distribution.
Finally, I present an integrated specification inference framework that interleaves inference and planning through active learning. The active learning models I developed enable the learner to select the learning modality with the highest expected utility and then identify and perform a task execution that would be most informative in refining its uncertainty. We demonstrate our framework by teaching a robot.
Created by Ankit Shah at Friday, July 30, 2021 at 3:28 PM.