Bilkent University
Department of Computer Engineering

---

M.S.THESIS PRESENTATION

 

PROACTIVITY IN ROBOTICS: TOWARDS LEARNING TO HELP WITHOUT BEING HELPED

 

Huzaifa Salahuddin
Master Student
(Supervisor: Asst.Prof.Özgür S.Öğüz)
Computer Engineering Department
Bilkent University

Abstract: With the rise of robots sharing spaces with humans, they should be capable of providing assistance without explicit human instructions. To this end, we propose proactive assistantship where the robotic agents are expected to act proactively based on inferred situational cues and observed environment such as placing a trash into trash can when encountering it on the floor, moving an apple from table into the fridge etc. We break down this problem into two sub-problems: first the agent must discover such objects of interest which demand attention and second, the agent should predict a correct set of actions to proactively assist humans in such tasks. We implement a novel context-aware scene encoder (CASE) which is capable of preserving environment context via spatial cues which is pivotal in deciding the correct proactivity task. Our subgoals prediction module uses graph neural networks (GNNs) to exploit the graph structure of situational context for accurate prediction of proactivity task. We validate our work through various experiments in our custom-designed simulation environments using Unity Gaming Engine. Our experiments show that utilizing graph-based architectures to solve this problem yields better performance as compared to other approaches due to the inherent graph structure of such proactive assistance problems.

 

DATE: August 29, Friday @ 08:30 Place: EA 409