Subteaming and Adaptive Formation Control for
Coordinated Multi-Robot Navigation

Author Names Omitted for Anonymous Review
Paper Supplementary Code Poster

Abstract

Coordinated multi-robot navigation is a fundamental capability for robots to operate as a team in diverse environments. During navigation, robot teams usually need to maintain specific formations, such as circular formations in order to protect human teammates positioned at the center. However, in complex scenarios such as narrow corridors, rigidly preserving the predefined team formations can become infeasible. Therefore, robot teams must have the capability for dynamically splitting into smaller subteams and adaptively controlling the subteams to navigate through these complex scenarios while preserving suitable formations. In this paper, we introduce a novel method for SubTeaming and Adaptive Formation (STAF) to enable coordinated multi-robot navigation in complex and challenging scenarios. STAF is built upon a unified hierarchical learning framework that incorporates three levels of robot learning: (1) high-level deep graph cut for team splitting, (2) intermediate-level graph learning for facilitating coordinated navigation among subteams, and (3) low-level policy learning for controlling individual mobile robots to reach their goal positions while avoiding collisions. In order to evaluate and validate STAF, we conducted extensive experiments in both indoor and outdoor environments using robotics simulations and physical robot teams. Experimental results have demonstrated that STAF enables the novel capability for subteaming and adaptive formation control, and achieves promising performance in coordinated multi-robot navigation through complex and challenging scenarios.

Motivation

Motivation image

Approach

Approach image

Overview of STAF, which integrates three levels of robot learning within a unified hierarchical learning framework to enable coordinated multi-robot navigation. It comprises:
(1) a high-level deep graph cut for team division,
(2) intermediate-level graph learning for formation adaptation,
(3) low-level RL-based policy learning for individual robot control in navigation and obstacle avoidance.

Evaluation on physical robot team running ROS2 in real indoor environments

We validate STAF involving real physical multi-robot teams, using differential-drive Limo robots equipped with caterpillar tracks. The indoor environments involve navigating constrained spaces, such as a narrow doorway in a hallway, a tight exit from an indoor area to a partially open outdoor space, and a slightly wider but still confined corridor.

Evaluation on physical robot team in outdoor environments with unstructured terrain

We conduct additional experiments in outdoor environments featuring unstructured terrain, including a narrow passage between two concrete security bollards, a forest-like environment with narrow corridors surrounded by scattered trees and obstacles, and a pathway with boundaries marked by two sticks blocking vehicle access.

Demonstration in Unity-based 3D multi-robot simulator in ROS1

We further use a high-fidelity simulator that integrates both the Unity3D engine for enhanced visual fidelity with ROS1 for multi-robot perception and control, which simulates outdoor field environments that include narrow pathways and bridges over creeks. In these experiments, we deploy ten differential-drive Warthog robots in circle and wedge formations, as well as nine Warthog robots in a line formation.

Comparison in Gazebo simulations in ROS1

We deploy ten Triton robots in the Gazebo simulations, and compare STAF with Leader and Follower method (L&F).