Cognitive AI Systems

[ICLR-2025] POGEMA stands for Partially-Observable Grid Environment for Multiple Agents. This is a grid-based environment that was specifically designed to be flexible, tunable and scalable. It can be tailored to a variety of PO-MAPF / MAPF settings.

[AAMAS-2026] MARL-GPT: This work tackles the challenge of a general-purpose model for multi-agent reinforcement learning. The proposed MARL-GPT uses offline RL on large-scale expert trajectories, leveraging a single GPT-based transformer encoder to train one model that performs competitively across diverse MARL environments (SMACv2, GRF and POGEMA)

[IROS-2025] MAPF-GPT-DDG is a scalable decentralized multi-agent pathfinding (MAPF) solver based on imitation learning. It builds upon MAPF-GPT by introducing a novel fine-tuning method called Delta Data Generation (DDG) — a reward-free active learning approach that identifies and corrects failure cases in the policy.

This is an umbrella repository that contains links and information about all the tools and algorithms related to the POGEMA Benchmark.

The Hierarchical Intrinsically Motivated Agent (HIMA) is an algorithm that is intended to exhibit an adaptive goal-directed behavior using neurophysiological models of the neocortex, basal ganglia, and thalamus.

[AAAI-2024] Follower: This study addresses the challenging problem of decentralized lifelong multi-agent pathfinding. The proposed Follower approach utilizes a combination of a planning algorithm for constructing a long-term plan and reinforcement learning for resolving local conflicts.

[AAAI-2024] MATS-LP addresses the challenging problem of decentralized lifelong multi-agent pathfinding. The proposed approach utilizes a combination of Monte Carlo Tree Search and reinforcement learning for resolving conflicts.