CODI addresses agent-quality imbalance in offline MARL by using LLMs and a compositional diffusion model. It decomposes team-level quality targets into fine-grained agent labels to generate balanced, high-quality trajectory segments for data augmentation. This approach effectively mitigates data imbalance and enables the learning of strong cooperative policies.

MAICC enhances coordination in MARL by using a decentralized memory retrieval system to provide agents with relevant context from past trajectories, enabling faster adaptation to new cooperative tasks without parameter updates.

CoPDT enables safe offline RL by prioritizing constraints in Decision Transformers, achieving 2× higher safety compliance than baselines across tasks.

Instead of discovering novel teammates only at the policy level, we utilize LLMs to propose novel coordination behaviors described in natural language, and then transform them into teammate policies, enhancing teammate diversity and interpretability, eventually learning agents with language comprehension ability and stronger collaboration skills.

We addresse the performance degradation of RL policies when state features (e.g., sensor data) evolve unpredictably by proposing Lapse, a method that reuses old policies by combining them with a state reconstruction model for vanished sensors and leverages past policy experience for offline training of new policies.

We propose a data augmentation technique for offline cooperative MARL, utlizing diffusion models to improve the quality of the datasets.

By compressing diverse best-response policies into a language-conditioned diffusion model, Haland efficiently aligns human preferences with AI behavior for seamless collaboration.

DiffMORL advances offline multi-objective RL with a diffusion-based planning framework, enhancing generalization via data mixup and outperforming baselines on OOD preferences.

We formulate domain calibration as a cooperative MARL problem to improve efficiency and fidelity.

We propose DASaR, which expands the trust region in sim-to-real RL by aligning simulator and real-world value functions through inverse dynamics-based relabeling of rewards and costs.

We study the problem of multi-objective reinforcement learning (MORL) with continually evolving learning objectives, and propose CORe3 to enable the MORL agent rapidly learn new objectives and avoid catastrophic forgetting about old objectives lacking reward signals.

We propose employing the large language models (LLMs) to develop an action plan (or equivalently, a convention) that effectively guides both human and AI for coordination.

We propose COSTA to deal with offline safe meta RL problems. We develope a cost-aware task inference module using contrastive learning to distinguish tasks based on safety constraints, and propose a safe in-distribution online adapation mechanism.

We review multi-agent cooperation from closed environment to open environment settings, and provide prospects for future development and research directions of cooperative MARL in open environments.

We propose Multi-agent Compatible Policy Learning (MACOP), where we adopt an agent-centered teammate generation process that gradually and efficiently generates diverse teammates covering the teammate policy space, and we use continual learning to train the ego agents to coordinate with them and acquire strong coordination ability.

We formulate Open Dec-POMDP and propose Fast teammate adaptation (Fastap) to enable controllable agents in a multi-agent system to fast adapt to the uncontrollable teammates, whose policy could be changed with one episode.

We formulate Limited Policy Adversary Dec-POMDP and propose ROMANCE to enable the trained agents to encounter diversified and strong auxiliary adversarial attacks during training, achieving high robustness under various policy perturbations.

We propose CroMAC to enable agents to obtain guaranteed lower bounds on state-action values to identify and choose the optimal action under a worst-case deviation when the received messages are perturbed.

We formulate the continual coordination framework and propose MACPro to enable agents to continually coordinate with each other when the dynamic of the training task and the multi-agent system itself changes over time.