Designing Versatile Robots that Learn, Assess, and Modify Tasks

Abstract

PhD Defense: Robots have the potential to improve our quality of life by performing dangerous tasks or tasks that are otherwise difficult for humans to perform across multiple domains. For robots to be versatile enough to autonomously act across multiple domains they need to learn new tasks online and improve over time. Furthermore, robots will need to learn how to interact with humans as they become more commonplace in human environments. Yet, robots are typically developed for specific domains with a set of pre-defined tasks and lack versatility and adaptability. The goal of this dissertation is to develop algorithms that enable a robot to be more versatile across multiple domains without needing to be reprogrammed. I present three algorithms which naturally interact with one another and allow a robot to adapt and become more versatile. Part I describes an interactive learning algorithm which allows a robot to learn multi-agent tasks through a single set of natural language instructions and zero or one demonstration. After learning a new task, a robot should assess its likelihood of completing the task successfully before it attempts to execute it. In Part II I present an algorithm that allows a robot to assess its expected performance for a task including, the probability of success, expected time-to-completion, and the action most likely to fail within a task. The human could then use the modification algorithm I present in Part III to alter how the robot executes a task. The task modification algorithm allows a human to instruct a modification to a task online through a single natural language instruction without needing to reprogram or reteach the task.

Join meeting in Cummings 610, or via Zoom:
https://tufts.zoom.us/j/8799487800
Password: see colloquium email
Dial-in not an options for this event.