A Framework for Real-World Multi-Robot Systems Running Decentralized GNN-Based Policies (2111.01777v2)

Abstract: GNNs are a paradigm-shifting neural architecture to facilitate the learning of complex multi-agent behaviors. Recent work has demonstrated remarkable performance in tasks such as flocking, multi-agent path planning and cooperative coverage. However, the policies derived through GNN-based learning schemes have not yet been deployed to the real-world on physical multi-robot systems. In this work, we present the design of a system that allows for fully decentralized execution of GNN-based policies. We create a framework based on ROS2 and elaborate its details in this paper. We demonstrate our framework on a case-study that requires tight coordination between robots, and present first-of-a-kind results that show successful real-world deployment of GNN-based policies on a decentralized multi-robot system relying on Adhoc communication. A video demonstration of this case-study, as well as the accompanying source code repository, can be found online. https://www.youtube.com/watch?v=COh-WLn4iO4 https://github.com/proroklab/ros2_multi_agent_passage https://github.com/proroklab/rl_multi_agent_passage

• The paper introduces a ROS2-based framework for deploying decentralized GNN policies, enabling robust multi-robot coordination.
• It evaluates communication strategies and network settings, revealing performance drops up to 36.4 percentage points in real-world scenarios compared to simulation.
• The framework’s decentralized approach enhances fault tolerance and scalability, paving the way for future research on bridging simulation-to-reality gaps.

A Framework for Decentralized Execution in Multi-Robot Systems via GNN-Based Policies

The paper presents a novel approach for deploying Graph Neural Network (GNN)-based policies in fully decentralized multi-robot systems by utilizing a framework built on the Robot Operating System 2 (ROS2). The authors explore the operational dynamics and performance of these policies in real-world scenarios, addressing the transition from simulation to real-world execution.

Overview

The introduction of decentralized GNN-based policies offers a promising method for handling complex multi-agent behaviors such as flocking, navigation, and control. GNNs provide an effective mechanism for communication and coordination among robots by leveraging latent messaging. This paper emphasizes the advantages of decentralized execution: removing single points of failure, enhancing fault tolerance, and improving scalability by distributing computation across multiple agents.

The framework proposed operates in four modes: centralized, offboard, onboard over infrastructure, and onboard over ad-hoc communication, each presenting different operational characteristics. These modes allow for varied execution sites and network configurations that address practical deployment challenges such as synchronous communication and scalability.

Technical Contributions

The paper makes several important contributions:

1. ROS2-Based Framework: A comprehensive software infrastructure using ROS2 that enables decentralized GNN implementation and testing in both simulated and real-world environments. This system supports automated resets and allows for continuous episode executions, facilitating large-scale experimental data collection.
2. Networking Solutions: Evaluation and optimization of communication strategies and network settings, including multicast versus unicast communication and 802.11 hardware retries, to enhance message delivery performance in decentralized settings. This aspect is crucial for maintaining efficient robotic coordination and minimizing latency impacts, as demonstrated in their extensive experiments.
3. Real-World Experiments: Deployment of GNN-based policies in a practical application where robots navigate through narrow passageways, illustrating the framework's applicability. The experiments show a decrease in performance transitioning from centralized simulation to decentralized real-world deployment due to communication and dynamic constraints.

Results

The experimental results highlight a performance drop in transitioning from simulation to real-world scenarios, with success rates decreasing by up to 36.4 percentage points. The centralized baseline outperformed the decentralized modes, with the real-world implementation showing increased makespans and a reduction in success rates due to communication latency and asynchronous execution challenges. Despite these challenges, the decentralized execution showed resilience and practical applicability under real-world conditions.

Implications and Future Work

The paper suggests significant implications for the future of decentralized multi-robot systems. The proposed framework paves the way for more robust deployments in complex and unstructured environments. By improving communication strategies and integrating mesh networking, the framework could better handle real-world variations in network conditions. Future research may focus on enhancing GNN robustness to domain shifts caused by communication delays and noisy environments, potentially bridging the simulation-to-reality gap further.

Overall, this work provides a foundational step in demonstrating decentralized GNN deployment, highlighting practical constraints and successes, and suggesting further enhancements to improve real-world applicability and performance.