HUMAN-GUIDED SUBGOAL LEARNING FOR SEQUENTIAL MANIPULATION IN NARROW AND CLUTTERED 2D MAZE-LIKE ENVIRONMENTS
Dilruba Sultan Haliloğlu
Master Student
(Supervisor: Asst.Prof.Özgür Salih Öğüz) Computer Engineering Department
Bilkent University
Abstract: Robotic agents should be able to perform sequential manipulation tasks since many real life tasks consist of interdependent sequential actions. Sequential manipulation in cluttered narrow spaces is a challenging issue in robotic planning because of the high dimensionality of the solution space and existence of local minima. One approach to solve such complex sequential manipulation tasks is to develop algorithms capable of decomposing these tasks into manageable subproblems. This work presents a framework that learns critical subgoals in a scene configuration from human data in 2D maze-like environments involving sequential object manipulation. We designed a data collection interface that allows participants to annotate subgoals for agents and objects in narrow and cluttered environments. We then use the collected data to train deep neural networks that predict subgoal target entity and possible subgoal positions as distributions in a given configuration. Finally, subgoals sampled from the predicted distributions are used to construct a final sequential manipulation plan to solve the task. The proposed pipeline generates a sequential manipulation plan for a diverse set of tasks with a success rate of 95.9%, demonstrating that using human-guided critical subgoal generation is a viable and promising approach.
DATE: September 09, Tuesday @ 10:00 Place: EA 409