CAMAR: Continuous Actions Multi-Agent Routing

Summary

This paper introduces CAMAR, a new test environment for multi-agent reinforcement learning where agents have to find paths in a continuous space, allowing for both working together and competing. It also proposes ways to test how well these learning systems are improving and lets researchers mix standard planning techniques with machine learning.

What's the problem?

Existing test environments for multi-agent reinforcement learning often don't have both continuous movement and complicated coordination or planning challenges. This makes it hard to test advanced learning strategies that need to handle smooth movements and complex teamwork or rivalry.

What's the solution?

The researchers created CAMAR, a benchmark specifically for multi-agent pathfinding with continuous actions. It supports cooperative and competitive scenarios, runs very fast, and allows for combining traditional planning methods like RRT with machine learning approaches. They also created a structured way to evaluate progress and provided tools for fair comparisons.

Why it matters?

This new benchmark is important because it provides a more realistic and challenging platform for developing and testing advanced multi-agent learning systems, especially those that need to navigate and coordinate in complex, continuous environments. This could lead to better AI for things like drone coordination or robotic teamwork.