Conditioning Multi-Agent Policies on Capabilities Enables Zero-Shot Generalization to New Robot Teams
Author(s)
Fu, Kevin
Advisor(s)
Editor(s)
Collections
Supplementary to:
Permanent Link
Abstract
In this work, we use MARL to solve a decentralized, partially-observable coordination problem for a team of heterogeneous robots, with a policy that can zero-shot generalize to different robot types and team compositions. To do so, we leverage the centralized training, decentralized execution (CTDE) paradigm to create a decentralized policy which can be copied to any number of robots, and condition a graph neural network (GNN) based policy on the capabilities of each robot to allow a single learned policy to zero-shot generalize to unseen robots and team compositions. We call this zero-shot generalizability adaptive teaming. Like prior work, our GNN facilitates learned communication of partial observations. Unlike prior work, we also demonstrate that communicating capability information among agents is vital for effective coordination in a heterogeneous multi-robot team. We find the constraint of needing to know the capabilities of each robot fairly realistic for real robotics use, as often real robots come with a specification sheet of known properties, such as maximum sensing range or top speed. We evaluate our policy's adaptive teaming ability on a cooperative robotics task we name Heterogeneous Sensor Network (HSN), a heterogeneous-robot variant of the common sensor coverage task. Our results demonstrate that our policy architecture, with awareness and communication of individual robot capabilities, outperforms existing agent-ID methods on adaptive teaming. Additionally, we provide results indicating that, allowing robots to communicate their own capabilities improves performance even without considering generalization as a necessary component. Finally, we share results of our policy coordinating real robots in the Robotarium, showing that our technique can be applied to real robots.
Sponsor
Date
Extent
Resource Type
Text
Resource Subtype
Undergraduate Research Option Thesis