GameChat: Multi-LLM Dialogue for Safe, Agile, and Socially Optimal Multi-Agent Navigation in Constrained Environments (2503.12333v1)

Abstract: Safe, agile, and socially compliant multi-robot navigation in cluttered and constrained environments remains a critical challenge. This is especially difficult with self-interested agents in decentralized settings, where there is no central authority to resolve conflicts induced by spatial symmetry. We address this challenge by proposing a novel approach, GameChat, which facilitates safe, agile, and deadlock-free navigation for both cooperative and self-interested agents. Key to our approach is the use of natural language communication to resolve conflicts, enabling agents to prioritize more urgent tasks and break spatial symmetry in a socially optimal manner. Our algorithm ensures subgame perfect equilibrium, preventing agents from deviating from agreed-upon behaviors and supporting cooperation. Furthermore, we guarantee safety through control barrier functions and preserve agility by minimizing disruptions to agents' planned trajectories. We evaluate GameChat in simulated environments with doorways and intersections. The results show that even in the worst case, GameChat reduces the time for all agents to reach their goals by over 35% from a naive baseline and by over 20% from SMG-CBF in the intersection scenario, while doubling the rate of ensuring the agent with a higher priority task reaches the goal first, from 50% (equivalent to random chance) to a 100% perfect performance at maximizing social welfare.

• The paper introduces GameChat, which combines multi-LLM dialogue with game theory and control barrier functions for safe, agile, and socially optimal multi-agent navigation.
• Evaluation showed GameChat with LLM communication achieved 100% correct priority assignments and significantly reduced time to goal compared to baselines.
• This approach ensures deadlock-free navigation and maximizes social welfare by prioritizing urgent tasks, demonstrating potential for complex real-world multi-robot systems.

Overview of GameChat: Multi-LLM Dialogue for Multi-Agent Navigation

The paper "GameChat: Multi-LLM Dialogue for Safe, Agile, and Socially Optimal Multi-Agent Navigation in Constrained Environments" (2503.12333) introduces a novel approach to address the challenges of decentralized multi-robot navigation in complex, constrained environments. The core problem lies in achieving safe, agile, and socially compliant navigation, especially when agents are self-interested and spatial symmetries can induce deadlocks. GameChat leverages natural language communication via LLMs to resolve conflicts, enabling agents to prioritize tasks based on urgency and break spatial symmetry in a socially optimal manner, with guarantees of subgame perfect equilibrium and safety through control barrier functions.

Key Contributions and Technical Details

GameChat's primary contributions revolve around integrating LLM-based communication with game-theoretic control strategies to achieve superior multi-agent navigation performance. The key contributions can be summarized as follows:

1. Subgame Perfect Equilibrium (SPE): GameChat employs a game-theoretic approach that guarantees SPE. This ensures that agents have no incentive to deviate from the agreed-upon behaviors during the navigation task.
2. Social Welfare Maximization: The algorithm is designed to maximize social welfare by prioritizing more urgent tasks. This effectively breaks spatial symmetry, a common cause of deadlocks in multi-agent systems.
3. Safety via Control Barrier Functions (CBFs): GameChat integrates CBFs to ensure safe trajectories and prevent collisions. This is crucial for real-world deployment where safety is paramount.
4. Agility Preservation: The approach minimizes disruptions to agents' planned trajectories, preserving smooth and agile movements. This ensures that agents can navigate efficiently without unnecessary detours or stops.

Technically, the approach involves several key components:

• Problem Formulation: The problem is mathematically framed as a modified Partially Observable Stochastic Game (POSG). The agents operate with single-integrator unicycle dynamics in a constrained environment. The environment consists of two agents that must navigate through a collision point.
• LLM-Based Communication: Agents use LLMs (specifically, gpt-4o-mini) to engage in natural language dialogue to determine task priorities. Each agent is assigned a task with varying urgency (e.g., going to the hospital, airport, or grocery store). The LLMs converse to decide which task is more urgent, facilitating socially optimal decision-making.
• Game-Theoretic Control Strategy: In scenarios where communication fails or during imminent collisions, the agents revert to a game-theoretic control strategy (Strategy 1). This involves agents assessing their time-to-collision (TTC) and assigning roles (leader/follower) to avoid collisions while minimizing disruptions.
• Safety and Agility Implementation: Safety is enforced through CBFs integrated into a Model Predictive Control (MPC) framework. Agility is maintained by minimizing deviations from the agents' desired trajectories.

Evaluation and Results

The effectiveness of GameChat was rigorously evaluated in simulated environments with doorways and intersections. Key metrics included the number of collisions, number of deadlocks, percentage of correct priority assignments, time to goal for the higher-priority agent, makespan, minimum velocity (as a measure of agility), and path deviation (as a measure of smoothness).

The results demonstrate significant improvements over baseline methods:

• GameChat (without LLM communication) outperformed baseline methods such as MPC-CBF and SMG-CBF in terms of deadlock resolution and agility (higher Min v and lower Makespan). MPC-CBF was shown to frequently lead to deadlocks.
• GameChat with LLM-based communication (both Pre-SMG Comm. and SMG Comm.) achieved 100% correct priority assignments, ensuring that the agent with the more urgent task is prioritized. The communicative method did have a slightly higher makespan than the noncommunicative method, and GameChat SMG Comm. had a slightly higher makespan than the Pre-Comm. variant.
• Specifically, GameChat reduces the time for all agents to reach their goals by over 35% from a naive baseline and by over 20% from SMG-CBF in the intersection scenario, while doubling the rate of ensuring the agent with a higher priority task reaches the goal first, from 50% (equivalent to random chance) to a 100% perfect performance at maximizing social welfare.

Addressing Multi-Agent Navigation Challenges

GameChat effectively addresses several critical challenges in multi-agent navigation:

• Conflict Resolution: Natural language communication enables agents to dynamically negotiate priorities based on their tasks, leading to more socially optimal and efficient conflict resolution.
• Deadlock Prevention: The game-theoretic control strategy ensures deadlock-free navigation by assigning roles and adjusting velocities to avoid collisions.
• Social Optimality: By prioritizing more urgent tasks, GameChat maximizes social welfare, ensuring that agents act in a socially responsible manner.
• Safety in Dynamic Environments: The integration of CBFs guarantees safety by preventing collisions, even in complex and constrained environments.

Conclusion

In conclusion, GameChat presents a robust and innovative solution for multi-agent navigation in constrained environments. By combining LLM-based communication with game-theoretic control and safety mechanisms, it achieves superior performance in terms of safety, agility, and social optimality. The experimental results validate the effectiveness of GameChat, demonstrating its potential for real-world deployment in various multi-robot applications.