Fall 2022 GRASP Seminar: Ankit Shah, Brown University, “Training Robots Like Apprentices”

*This is a HYBRID Event with in-person attendance in Levine 307 and virtual attendance via Zoom.

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, or in other words robotic apprentices that learn from experts will drive the next wave of robotics adoption.

In this talk, I present a multi-modal Bayesian framework for teaching a robot learner to identify the teacher’s intended task from natural teaching modalities such as demonstrations, and acceptability assessments of the execution. The framework centers on using formal languages such as LTL to model the task specification, and using probabilistic reasoning to reason about the ambiguity in natural teaching modalities. Utilizing the Bayesian framework, we can teach the robot the task specifications from demonstrations, and the robot models its updated belief over specifications as a distribution of logical formulas. We propose Planning with Uncertain Specifications (PUnS), a novel framework to reason about the uncertainty of task specifications while computing the robot policy. We also demonstrate how using formal languages along with active learning can help the robot refine its belief efficiently. Finally we demonstrate how the temporal abstractions afforded by temporal logics in particular can help the robot learn to reuse policies from one task to accomplish other closely related tasks without any additional learning.