MAPF-GPT – a deep learning-based model for solving MAPF problems (github.com)

Researchers have announced the release of MAPF-GPT, a cutting-edge deep learning model specifically designed for multi-agent pathfinding (MAPF) problems. The model is built using a substantial dataset comprising 1 billion tensor-action pairs, with 90% derived from maze maps, to facilitate efficient training and validation. With scripts for training, evaluation, and data generation available, users can easily launch MAPF-GPT using pre-built configurations and run various scenarios from the POGEMA benchmark, providing insights into performance metrics such as success rates and efficiency through visual outputs. The significance of MAPF-GPT lies in its potential to enhance multi-agent systems across various applications, including robotics and autonomous vehicle navigation. By leveraging the LaCAM approach for expert data sourcing and utilizing robust training mechanisms that support multiple GPUs, the model achieves scalability and effectiveness. Users can select from different model sizes, allowing them to tailor the solution to their hardware capabilities—especially helpful for environments like Apple Silicon. The flexibility in dataset generation and fine-tuning capabilities further increases the utility of MAPF-GPT for researchers and practitioners in the AI/ML community, enabling more adaptable and powerful solutions for complex pathfinding challenges.