Event
Ph.D. Research Proposal Exam: Ariana Y. Spalter
Tuesday, May 28, 2024
11:00 a.m.
AVW 2168
Maria Hoo
301 405 3681
mch@umd.edu
ANNOUNCEMENT: Ph.D. Research Proposal Exam
Name: Ariana Y. Spalter
Committee:
Professor John S. Baras (Chair)
Professor Dinesh Manocha
Professor Eyad Abed
Date/time: Tuesday May 28, 2024 at 11:00 am
Location: AVW 2168
Title: Hard Constraint Checking and Enforcement for Operating in Uncertain, Shared Human-Robot Workspaces
Abstract:
We are in an aging society where the number of older adults is rapidly becoming disproportionately larger than the number of younger adults in the world. This means that soon there will be a large gap between the number of those requiring care and the number of those able to provide care. Assistive devices will therefore be needed to help fill this gap and extend the ability of people to live independently. Specifically, for household activities which require use of object manipulation, multi-step task completion, or other physical interactions, the use of assistive robots will be key. We ground our problems within the domain of a home kitchen environment due to its high variability, multitude of tasks, and necessity in performing daily life activities. Meal preparation tasks require complex cognitive reasoning capabilities and has been found to be an area that large numbers of older adults start having trouble with as they age. We therefore ground our problems in this domain where the robotic tasks we focus on help augment the skills of the user through acting as a kitchen assistant in tasks that have variability in available items, user preferences, and frequency of task interruptions. However, before an assistive robot can be used in a shared environment with a human user, it must first be able to deal with handling uncertainties which may occur during deployment.
We propose a robotic system capable of taking pre-learned motions and ensuring new hard constraints defined during run-time can be guaranteed so that the system can operate in uncertain, real-world environments. The system considerations can be classified at two levels of operation, low-level collision avoidance and high-level task planning for user preferences and safety constraints. For the low-level considerations, we propose a method using a trajectory tracking controller to guarantee collision avoidance of obstacles unknown to a pre-learned motion during its training phase. Further, we plan to develop a way to look ahead a few steps into the future to better plan how to avoid obstacles. In doing so, we will extend upon methods doing one-step checks to plan for safer long term motions well in advance of possible collisions. For the high-level considerations, we plan to cover new user preferences and safety constraints which can be added to a task plan during the execution of the task. Such high-level preferences will be represented using Non-Monotonic Metric Temporal Logic. This means that the high-level logic formulations will be able to cover not only the ordering of tasks but also exceptions and time intervals restricting when the task should take place. We then aim to scale these problems to a shared environment with a human. This last problem will aim to cover low-level dynamic obstacle avoidance and high-level user intent prediction. Dynamic obstacle avoidance will be explored using reachability analysis under the speed and separation monitoring protocols. Intent prediction will be used to predict the most likely next goal of the user to block off overlapping robot goals.
To verify and validate this proposed robotic system, we consider two real-world scenarios. The first focuses on a robot kitchen assistant moving a knife across a cluttered shared tabletop environment towards a goal position to validate low-level guarantees for collision avoidance and joint limits. This lends itself to different objects interrupting the workflow where looking ahead could help improve performance. The second focuses on a robot kitchen assistant helping with a lunchbox packing task to validate high-level guarantees for safety constraints and ad hoc user preferences during run-time. This could support re-ordering, adding, or taking away sub-tasks in the plan based on new user constraints like allergies to certain food options. These two tasks can be linked in that they both aim to address key factors of real-time operation of a complex robotic manipulation system such as safety, changing user preferences, and the ability to deal with uncertainty.
