Decentralized Collective Transport of Unknown Complex Objects Using Global State Prediction​

Poindexter, Julian

Etd

Decentralized Collective Transport of Unknown Complex Objects Using Global State Prediction

公开 Deposited

Reinforcement Learning (RL) has been shown to be successful in the control of multi-agent systems. While RL is becoming more widely used in multi-agent control, policies can struggle with convergence in decentralized systems due to the dynamics of the environment changing over time. To study this problem, in this thesis we focus on a complex application scenario — collective transport. In collective transport, a swarm of robots must coordinate to carry a heavy payload to a target location avoiding obstacles along the way. Currently, the solutions to collective transport based on RL assume prior knowledge of the payload, and particularly its shape and center of mass. This assumption is unrealistic in many real-world applications. We examined the use of a novel approach to RL, Global State Prediction (GSP), which enables the individual robots in the swarm to make predictions about the future state of the swarm using only locally communicated information. Our experiments show that using GSP yields viable transport behaviors faster than RL without GSP. The behaviors found by GSP are also more efficient than those found without it.

Creator