Bilkent University
Department of Computer Engineering
CS 590/690 SEMINAR
Way-TU: Learning Waypoint Representations for Tool Usage
Ece Kunduracıoğlu
Master Student
(Supervisor: Asst.Prof.Özgür Salih Öğüz)
Computer Engineering Department
Bilkent University
Abstract: Integrating intelligent robots into daily life depends on their ability to select and effectively manipulate tools, enabling them to perform tasks that would be unfeasible without tool assistance. Sparse keypoint representations have proven effective in guiding robots to manipulate tools for object interaction. In this work, we present Way-Tu, a novel framework that combines tool selection and manipulation skills within a unified network. Way-Tu uses waypoints, defined as keypoints with orientations, to generate auxiliary points that improve the optimization-based motion planning process, enhancing robotic capabilities in tool-based tasks. We demonstrated the capabilities of our framework across three tasks: reaching, pushing (mini-golf), and lifting. Using six distinct primitive tool types, we show that our network can generate waypoints to solve tasks with various tool geometries. While these initial results highlight the promise of our approach, we are actively enhancing the network to expand its performance across a broader range of tasks and more dynamic environments.
DATE: November 25, Monday @ 16:10 Place: EA 502