Decentralized Collective Transport of Unknown Complex Objects Using Global State Prediction
Abstract: Reinforcement Learning (RL) has been shown to be successful in the control of multi-agent systems. In complex coordination tasks like collective transport, RL can be incredibly useful in finding optimal behavior. 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. Currently, many solutions to this issue in regards to collective transport rely on the use of prior knowledge about the payload such as its shape or center of mass. This makes it unrealistic to apply to real-world applications. We examined the use of Global State Prediction (GSP) to make predictions about the future state of the group using only locally communicated information. We implement GSP in a decentralized collective transport scenario to move objects of unknown shape and mass. GSP policy displays the ability to make accurate predictions about the trajectory of objects without any explicit knowledge of its shape and mass.
Advisor: Professor Carlo Pinciroli
Committee: Professor Kevin Leahy, Professor Constantinos Chamzas
Also, available to attend via Zoom: https://wpi.zoom.us/j/99138927367