Skip to main content

RBE PhD Speaking Qualifier: Joshua Bloom | Multi-Agent Reinforcement Learning under Physical Constraints

robotics_banner.jpg

Various images of robots at Robotics Engineering WPI alt
WPI Robotics Engineering
Tuesday, December 07, 2021
3:30 pm to 4:30 pm

Robotics Engineering Department

Ph.D. Speaking Qualifier

 

Joshua Bloom

 Multi-Agent Reinforcement Learning under Physical Constraints

 

Tuesday, December 7, 2021

3:30 Pm – 4:30 Pm

Virtual | Please email Joshua Bloom (jdbloom@wpi.edu) for the zoom link

 

Abstract:  In this paper, we present an approach to apply decentralized reinforcement learning to a scenario in which a swarm of robots must perform collective transport. A key feature in our work is that we assume the robots to be minimalistic. The robots can sense the target location and immediate obstacles, but lack the means to communicate explicitly through, e.g., message-passing. What makes this scenario compelling is the fact that the robots are constrained to be physically attached to the object to transport. Due to decentralization and to the absence of explicit communication, the robot must learn to cooperate by pushing and pulling the object in a coordinate manner, while avoiding obstacles. We compare two well-known RL approaches, DQN and DDQN, augmented with a suitably defined training curriculum in trials involving different kinds of obstacles and failures. Our results suggest that DDQN outperforms DQN in learning effective collective transport behaviors.

Ph.D. Advisor:                      

Professor Carlo Pinciroli, Worcester Polytechnic Institute (WPI)

Ph.D. Qualifier Review Committee:

Professor Cagdas Onal, Worcester Polytechnic Institute (WPI)

Professor Berk Calli, Worcester Polytechnic Institute (WPI)

DEPARTMENT(S):