Metis-Core New Release: 4 AI Drones Collaborate in Search-and-Rescue Simulation

Open-Source Reinforcement Learning Framework Metis-Core Releases SARSwarm Example

Recently, the developer released Metis-Core v0.4.2-Alpha, a reinforcement learning framework written in C++, along with an example project called SARSwarm: a simulation of 4 radio-coordinated drones searching for shipwreck survivors. The example demonstrates collaboration and communication modeling in multi-agent systems for search-and-rescue tasks, providing reinforcement learning enthusiasts with a runnable sandbox environment.

According to the developer, in SARSwarm, 4 drones take off from a mother ship with the goal of locating a person in the water. They need to learn to coordinate with each other and avoid repeatedly scanning areas already covered by other drones. To make the simulation more realistic, the project models radio communication: once a connection is lost, the drones stop sharing maps with each other. As time passes, previously scanned areas become "colder," signaling to the drones that the person in the water may have re-entered that area due to random drifting from ocean currents.

Multi-Agent Reinforcement Learning (MARL) is a branch of reinforcement learning that studies how multiple agents in the same environment learn cooperative or competitive strategies. Unlike single-agent scenarios, MARL faces challenges such as environmental non-stationarity (other agents are also learning, causing the environment dynamics to change), credit assignment (how to attribute team rewards to individuals), and communication constraints. Examples like SARSwarm typically adopt the Centralized Training with Decentralized Execution (CTDE) paradigm, where global information is allowed during training, but at execution time each agent relies only on local observations and limited communication.

Multi-Agent Collaboration and Communication Modeling

The core challenge of SARSwarm lies in coordination among multiple agents. Each drone has independent perception and decision-making capabilities, but must share information through communication to efficiently cover the search area. The developer deliberately introduced a communication disruption mechanism, forcing agents to make decisions under incomplete information. This design brings the simulation closer to real search-and-rescue scenarios, where maritime communication is often unstable.

The setting of "colder" areas adds further dynamism. Ocean currents cause the position of the person in the water to change randomly, and the drones need to re-examine previously scanned areas rather than simply flying along fixed routes. This requires agents to have a certain degree of memory and reasoning ability, enabling them to adjust search strategies as time progresses.

In SARSwarm, radio communication disruption prevents map sharing, which falls under the problem of collaboration under communication constraints. Common modeling approaches include: learning communication actions as part of the agent's policy (learning "when to speak and to whom"), or adopting a fixed communication topology while introducing random packet loss. After a connection is lost, each drone can only rely on its own historical observations to maintain a local map, requiring the introduction of belief states or recurrent neural networks to estimate the state of unobserved areas—consistent with the formulation of Partially Observable Markov Decision Processes (POMDPs).

Technical Implementation and Community Feedback

Metis-Core is implemented in C++, with a focus on performance and scalability. The graphical interface of the SARSwarm example is developed based on the Raylib library, providing intuitive visualization. The developer encourages users to report compilation, crash, and other issues through the GitHub discussion forum, and has stated they will actively assist in fixing them.

The project is currently in the Alpha stage (v0.4.2), meaning that features and interfaces may still be unstable. However, for researchers studying multi-agent reinforcement learning, cooperative strategies under communication constraints, and search-and-rescue path planning, SARSwarm provides a valuable experimental platform. The developer also provides a tutorial repository, Tutorial-MultiHead-SARSwarm, to help users get started quickly.

Raylib is a lightweight open-source C language game development library that encapsulates the low-level details of OpenGL, making it suitable for quickly building 2D/3D visualization interfaces. In reinforcement learning projects, Raylib is often used to render agent positions, coverage areas, and communication links in real time without introducing a heavyweight game engine. Metis-Core's choice of Raylib as the graphical interface foundation for SARSwarm provides low-overhead visual debugging while maintaining a high-performance C++ training loop.

Potential Applications and Outlook

Although SARSwarm is a simulation example, the technical ideas behind it can be extended to real-world scenarios. Multi-drone collaborative search and rescue, task planning in communication-constrained environments, and dynamic target search are all hot topics in the reinforcement learning field. As an open-source framework, Metis-Core is expected to attract more developers to contribute algorithms and examples, driving the development of multi-agent systems.

The developer welcomes community suggestions and feedback to jointly improve the project. With subsequent version iterations, Metis-Core may support more complex scenarios and algorithms, making it worthy of continued attention.