[Paper Review] MRNAV: Multi-Robot Aware Planning and Control Stack for Collision and Deadlock-free Navigation in Cluttered Environments

Baskın Şenbaşlar, Pilar Luiz+2 more|arXiv (Cornell University)|Aug 25, 2023

Robotic Path Planning AlgorithmsComputer Science3 citations

TL;DR

MRNAV proposes a hierarchical multi-robot navigation stack using long, medium, and short horizon decision-making modules to achieve collision- and deadlock-free navigation in cluttered environments with static and dynamic obstacles. The framework ensures real-time executability, dynamic feasibility, and robustness to communication issues, demonstrating eight hours of uninterrupted, collision-free simulation with eight quadrotors.

ABSTRACT

Multi-robot collision-free and deadlock-free navigation in cluttered environments with static and dynamic obstacles is a fundamental problem for many applications. We introduce MRNAV, a framework for planning and control to effectively navigate in such environments. Our design utilizes short, medium, and long horizon decision making modules with qualitatively different properties, and defines the responsibilities of them. The decision making modules complement each other and provide the effective navigation capability. MRNAV is the first hierarchical approach combining these three levels of decision making modules and explicitly defining their responsibilities. We implement our design for simulated multi-quadrotor flight. In our evaluations, we show that all three modules are required for effective navigation in diverse situations. We show the long-term executability of our approach in an eight hour long wall time (six hour long simulation time) uninterrupted simulation without collisions or deadlocks.

Motivation & Objective

• To address the challenge of safe, deadlock-free navigation for multi-robot teams in complex, cluttered environments with static and dynamic obstacles.
• To design a hierarchical planning and control stack that integrates qualitatively different decision-making modules rather than treating them as alternatives.
• To ensure real-time executability, dynamic feasibility, and resilience to communication imperfections such as message drops and delays.
• To demonstrate long-term system stability and effectiveness across diverse environments, including forests and mazes with dynamic obstacles.
• To validate that all three decision-making modules—long, medium, and short horizon—are jointly necessary for robust performance in all scenarios.

Proposed method

• The framework employs three distinct decision-making modules: long-horizon (LH-NDM) for global path planning with completeness guarantees, medium-horizon (MH-NDM) for local trajectory adjustment and deadlock resolution, and short-horizon (SH-NDM) for real-time, reactive collision avoidance.
• LH-NDM generates globally feasible, dynamically feasible trajectories using optimization-based planners like RRT* or A*, ensuring goal reachability when a solution exists.
• MH-NDM provides intermediate-level reasoning to resolve local deadlocks by adjusting trajectories while respecting dynamic constraints and maintaining system liveness.
• SH-NDM enforces dynamic feasibility and safety at the action level, ensuring every executed action is collision-free, even when higher-level plans fail or diverge.
• The modules are tightly integrated via well-defined interfaces, with SH-NDM acting as a safety layer that enables reactivity and recovery from deviations.
• The system is implemented and evaluated in a simulated multi-quadrotor environment with motor-speed control, supporting dynamic obstacle prediction and communication imperfections.

Experimental results

Research questions

• RQ1 Can a hierarchical integration of long, medium, and short horizon decision-making modules achieve collision- and deadlock-free navigation in diverse cluttered environments with dynamic obstacles?
• RQ2 Is the combination of all three modules necessary, or can any single module suffice in all scenarios?
• RQ3 How does the system maintain long-term stability and safety under communication delays, message drops, and dynamic obstacle unpredictability?
• RQ4 What role does medium-horizon planning play in resolving local deadlocks when long-horizon plans fail due to dynamic obstacles?
• RQ5 To what extent does the SH-NDM module ensure dynamic feasibility and safety during execution, especially when trajectories diverge from reference paths?

Key findings

• The combination of long, medium, and short horizon decision-making modules (LH+MH+SH) is the only configuration that achieves collision- and deadlock-free navigation across all tested environments, including forests and mazes with dynamic obstacles.
• In the absence of dynamic obstacles, the LH+SH combination is effective, but fails when dynamic obstacles are introduced due to inability to handle dynamic collision avoidance without medium-horizon planning.
• The SH-NDM module is essential for safety, as all successful configurations include it, enabling reactive collision avoidance and recovery from trajectory divergence.
• The system successfully executed an uninterrupted eight-hour simulation (six hours of simulated time) with eight quadrotors navigating in a forest-like environment containing 300 dynamic obstacles, with zero collisions or deadlocks.
• Adding the MH-NDM module increases navigation duration when the LH+SH combination is already effective, due to imposed dynamic limits for real-time executability, but is critical for handling dynamic obstacles.
• Global deadlock resolution requires global reasoning, which only the LH-NDM module provides, making it indispensable in complex environments like mazes regardless of obstacle dynamics.