Multi-Agent Reinforcement Learning

Progress in adversarial multi-agent reinforcement learning (MARL) for robotics has been hampered by a lack of shared, extensible infrastructure that supports heterogeneous agent morphologies in high-fidelity physics simulation. Existing frameworks either focus on cooperative tasks, rely on simplified physics engines, or provide isolated implementations that are difficult to extend. We present HARL-A, an open-source, actively maintained framework built on IsaacLab that enables scalable training and benchmarking of adversarial policies across morphologically diverse robot teams with any number of teams and any mix of robot morphologies per team. HARL-A extends the HARL algorithm library and IsaacLab with adversarial multi-agent support and contributes three components: (1) a modular software architecture that reduces the engineering overhead of defining new heterogeneous adversarial environments, (2) a suite of three benchmark environments---Sumo, Soccer, and 3D Galaga---spanning contact-rich pushing, ball-skill competition, and pursuit/evasion, (3) over ten pretrained policies spanning homogeneous and heterogeneous team configurations, released publicly on Hugging Face to enable immediate exploration of adversarial learning dynamics without retraining from scratch. We demonstrate the framework across multiple competitive scenarios, showing that it reliably produces learned adversarial policies and emergent role specialization. All code environments, trained policies, and documentation are openly available at https://github.com/DIRECTLab/IsaacLab-HARL.

Multi-agent games in dynamic nonlinear settings are challenging due to the time-varying interactions among the agents and the non-stationarity of the (potential) Nash equilibria. In this paper we consider model-free games, where agent transitions and costs are observed without knowledge of the transition and cost functions that generate them. We propose a novel distributed policy structure that follows the communication constraints in multi-team games, with multiple agents per team, and learned through policy gradients. Our formulation is inspired by the structure of distributed policies in linear quadratic games, which take the form of time-varying linear feedback gains. In the nonlinear case, we model the policies as nonlinear feedback gains, parameterized by self-attention layers to account for the time-varying multi-agent communication topology. We demonstrate that our approach achieves strong performance in several settings, including distributed linear and nonlinear regulation, and simulated and real multi-robot pursuit-and-evasion games.

We propose and study distributionally robust Markov games (DR-MGs) with the average-reward criterion as a crucial framework for multi-agent decision-making under model mismatches and over extended horizons. Under a standard irreducible assumption, we first derive a correspondence between the optimal policies and the solutions of the robust Bellman equation, based on which we further show the existence of a stationary Nash Equilibrium (NE) of the game. We further study DR-MGs under a more general weakly communicating setting. We construct a set-valued map based on the constant-gain optimal robust Bellman operator and show that its value is a subset of the best-response policies. We further prove that this map admits a fixed point, which implies the existence of NE. We then design two algorithms, Robust Nash-Iteration and robust TD Descent, with provably convergent guarantees. Finally, we show that the NE under average-reward can be approximated by the ones for the discounted DR-MGs as the discount factor approaches one. Our studies provide a comprehensive theoretical and algorithmic foundation for decision-making in complex, uncertain, and long-running multi-player environments.

Decentralized Multi-Agent Reinforcement Learning (MARL) methods allow for learning scalable multi-agent policies, but suffer from partial observability and induced non-stationarity. These challenges can be addressed by introducing mechanisms that facilitate coordination and high-level planning. Specifically, coordination and temporal abstraction can be achieved through communication (e.g., message passing) and Hierarchical Reinforcement Learning (HRL) approaches to decision-making. However, optimization issues limit the applicability of hierarchical policies to multi-agent systems. As such, the combination of these approaches has not been fully explored. To fill this void, we propose a novel and effective methodology for learning multi-agent hierarchies of message-passing policies. We adopt the feudal HRL framework and rely on a hierarchical graph structure for planning and coordination among agents. Agents at lower levels in the hierarchy receive goals from the upper levels and exchange messages with neighboring agents at the same level. To learn hierarchical multi-agent policies, we design a novel reward-assignment method based on training the lower-level policies to maximize the advantage function associated with the upper levels. Results on relevant benchmarks show that our method performs favorably compared to the state of the art.

Long-horizon agentic systems become long-context systems by accumulation rather than through a single long document, since every step reassembles a prompt from role instructions, tool and sub-agent descriptions, recalled memories, dialogue history, and previous tool outputs, separately for every agent. This paper studies how such accumulated context can be trained, evaluated, and managed in hierarchical multi-agent assistants. We present HiMA-Ecom, a hierarchical multi-agent benchmark for e-commerce assistants with 22.8K instances, of which 17.7K carry memory entries in context, pairing agent-specific supervised fine-tuning data with system-level supervision for end-to-end multi-agent reinforcement learning. Building upon it, HiMA-R1 keeps the context budget of a trajectory under control through Variance-Reduction Group Relative Policy Optimization (VR-GRPO), which samples along an initial trajectory so that sampling cost grows additively with trajectory length and updates the agents with the largest reward variance, together with an efficiency reward on trajectory length and a memory evolution mechanism that repurposes GRPO rewards as cost-free signals for what stays in context. Agents built upon 3B and 7B open-source models match DeepSeek-R1 and surpass DeepSeek-V3 by 6.0 points on average, while reward-driven memory reduces the update steps needed for convergence by roughly 55%.

Conventional large language model (LLM) safety alignment relies on a reactive, disjoint loop: attackers exploit a static model, then defenders patch exposed vulnerabilities. This sequential setup leads to attackers overfitting obsolete exploits while defenders perpetually lag behind emerging threats. To address this, we introduce Self-RedTeam, the first fully online self-play multi-agent reinforcement learning (MARL) algorithm that continuously co-evolves attacker and defender for robust safety alignment. A single policy self-plays as both attacker and defender, generating adversarial prompts and defending against them, with a reward model adjudicating outcomes. Each role uses hidden chain-of-thought for strategic planning. Grounded in two-player zero-sum game theory, we establish a theoretical safety guarantee: if the game converges to Nash Equilibrium, the defender produces safe responses against any adversarial input. Empirically, Self-RedTeam generalizes across five models from the Llama and Qwen families, uncovering more diverse attacks (+17.80% SBERT) and improving safety of RLHF-trained models by up to 95% across 14 benchmarks. Our work motivates a shift from reactive patching to proactive co-evolution, enabling LLM safety self-improvement via online self-play MARL. Link to code: https://github.com/mickelliu/selfplay-redteaming

Multi-agent reinforcement learning (MARL) methods have achieved state-of-the-art results on a range of multi-agent tasks. Yet, MARL algorithms typically require significantly more environment interactions than their single-agent counterparts to converge, a problem exacerbated by the difficulty in exploring over a large joint action space and the high variance intrinsic to MARL environments. To tackle these issues, we propose a novel algorithm that combines a decomposed centralized critic with decentralized ensemble learning, incorporating several key contributions. The main component in our scheme is a selective exploration method that leverages ensemble kurtosis. We extend the global decomposed critic with a diversity-regularized ensemble of individual critics and utilize its excess kurtosis to guide exploration toward high-uncertainty states and actions. To improve sample efficiency, we train the centralized critic with a novel truncated variation of the TD($λ$) algorithm, enabling efficient off-policy learning with reduced variance. On the actor side, our suggested algorithm adapts the mixed samples approach to MARL, mixing on-policy and off-policy loss functions for training the actors. This approach balances between stability and efficiency and outperforms purely off-policy learning. The evaluation shows our method outperforms state-of-the-art baselines on standard MARL benchmarks, including a variety of SMAC II maps.

While multi-agent reinforcement learning (MARL) has been proven effective across both collaborative and competitive tasks, existing algorithms often struggle to scale to large populations of agents. Recent advancements in mean-field (MF) theory provide scalable solutions by approximating population interactions as a continuum, yet most existing frameworks focus exclusively on either fully cooperative or purely competitive settings. To bridge this gap, we introduce MF-MAPPO, a mean-field extension of PPO designed for zero-sum team games that integrate intra-team cooperation with inter-team competition. MF-MAPPO employs a shared actor and a minimally informed critic per team and is trained directly on finite-population simulators, thereby enabling deployment to realistic scenarios with thousands of agents. We further show that MF-MAPPO naturally extends to partially observable settings through a simple gradient-regularized training scheme. Our evaluation utilizes large-scale benchmark scenarios using our own testing simulation platform for MF team games (MFEnv), including offense-defense battlefield tasks as well as variants of population-based rock-paper-scissors games that admit analytical solutions, for benchmarking. Across these benchmarks, MF-MAPPO outperforms existing methods and exhibits complex, heterogeneous behaviors, demonstrating the effectiveness of combining mean-field theory and MARL techniques at scale.

Large Language Model (LLM)-based Multi-Agent Systems (LaMAS) offer a promising approach to collaborative tasks, but applying reinforcement learning to these systems remains challenging because classical multi-agent reinforcement learning assumptions do not fully capture their dependency structures, heterogeneous roles, and autoregressive outputs. This article provides a tutorial-style overview of the transition from reinforcement learning to reinforcement fine-tuning in single-agent and multi-agent settings, highlighting the differences between classical MARL and LLM-based multi-agent optimization. We introduce Multi-Agent Reinforcement Fine-Tuning (MARFT) and formalize LaMAS optimization as a Flexible Markov Game (Flex-MG). The MARFT framework combines predecessor-conditioned policy optimization, a centralized critic, parameter-efficient adapters, and token-level credit assignment. Our token-level analysis clarifies how within-action discounting affects returns and establishes consistency with action-level returns when the token discount is one. Experiments on fixed two- to four-agent mathematical and coding workflows compare MARFT with and without policy parameter sharing against IPPO. Both variants report higher mean evaluation scores in all 24 benchmark-team-size settings, while the variant without parameter sharing shows the clearest advantage in final training reward. Role-specification ablations examine how prescribed roles affect training. We conclude with application perspectives and open challenges, including dynamic workflows, sample efficiency, and system-level reliability. Code: https://github.com/jwliao-ai/MARFT.

Recent advancements in multi-agent reinforcement learning (MARL) have opened up vast application prospects, such as swarm control of drones, collaborative manipulation by robotic arms, and multi-target encirclement. However, potential security threats during the MARL deployment need more attention and thorough investigation. Recent research reveals that attackers can rapidly exploit the victim's vulnerabilities, generating adversarial policies that result in the failure of specific tasks. For instance, reducing the winning rate of a superhuman-level Go AI to around 20%. Existing studies predominantly focus on two-player competitive environments, assuming attackers possess complete global state observation. In this study, we unveil, for the first time, the capability of attackers to generate adversarial policies even when restricted to partial observations of the victims in multi-agent competitive environments. Specifically, we propose a novel black-box attack (SUB-PLAY) that incorporates the concept of constructing multiple subgames to mitigate the impact of partial observability and suggests sharing transitions among subpolicies to improve attackers' exploitative ability. Extensive evaluations demonstrate the effectiveness of SUB-PLAY under three typical partial observability limitations. Visualization results indicate that adversarial policies induce significantly different activations of the victims' policy networks. Furthermore, we evaluate three potential defenses aimed at exploring ways to mitigate security threats posed by adversarial policies, providing constructive recommendations for deploying MARL in competitive environments.

Benchmarks are crucial in the development of machine learning algorithms, with available environments significantly influencing reinforcement learning (RL) research. Traditionally, RL environments run on the CPU, which limits their scalability with typical academic compute. However, recent advancements in JAX have enabled the wider use of hardware acceleration, enabling massively parallel RL training pipelines and environments. While this has been successfully applied to single-agent RL, it has not yet been widely adopted for multi-agent scenarios. In this paper, we present JaxMARL, the first open-source, Python-based library that combines GPU-enabled efficiency with support for a large number of commonly used MARL environments and popular baseline algorithms. Our experiments show that, in terms of wall clock time, our JAX-based training pipeline is around 14 times faster than existing approaches, and up to 12500x when multiple training runs are vectorized. This enables efficient and thorough evaluations, potentially alleviating the evaluation crisis in the field. We also introduce and benchmark SMAX, a JAX-based approximate reimplementation of the popular StarCraft Multi-Agent Challenge, which removes the need to run the StarCraft II game engine. This not only enables GPU acceleration, but also provides a more flexible MARL environment, unlocking the potential for self-play, meta-learning, and other future applications in MARL. The code is available at https://github.com/flairox/jaxmarl.

Traditional methods plan feasible paths for multiple agents in the stochastic environment. However, the methods' iterations with the changes in the environment result in computation complexities, especially for the decentralized agents without a centralized planner. Although reinforcement learning provides a plausible solution because of the generalization for different environments, it struggles with enormous agent-environment interactions in training. Here, we propose a novel centralized training with decentralized execution method based on multi-agent reinforcement learning, which is improved based on the idea of model predictive control. In our approach, agents communicate only with the centralized planner to make decentralized decisions online in the stochastic environment. Furthermore, considering the communication constraint with the centralized planner, each agent plans feasible paths through the extended observation, which combines information on neighboring agents based on the distance-weighted mean field approach. Inspired by the rolling optimization approach of model predictive control, we conduct multi-step value convergence in multi-agent reinforcement learning to enhance the training efficiency, which reduces the expensive interactions in convergence. Experiment results in both comparison, ablation, and real-robot studies validate the effectiveness and generalization performance of our method.

Many real-world domains--including e-commerce platform design, security planning, and multi-agent coordination--feature leader-follower problems where one decision-maker commits to a policy and others react strategically. We develop a reinforcement learning framework for such interactions in sequential environments with partial observations and multiple followers. Followers may adapt through no-regret learning or reinforcement learning, potentially departing from equilibrium behavior. The framework embeds follower adaptation into the leader's environment to construct a single-agent partially observable Markov decision process--the Stackelberg POMDP. For policy-interactive response algorithms, which access the leader's policy through queries, we prove that an optimal policy based only on the leader's game history yields an optimal commitment under the specified response procedure. We use proximal policy optimization with a centralized critic and train contextual meta-followers to respond across leader policies. In indirect mechanism design, mechanisms using buyer messages achieve higher social welfare than optimal standard sequential price mechanisms across all tested type counts, with responses certified as approximate Bayesian coarse correlated equilibria. In platform design, learned display rules increase mean consumer surplus by 8.4% over an optimized fixed price cap while accommodating hidden seller costs. In Atari bilateral trade, meta-learned follower responses support joint learning of visual gameplay and economic decisions; assigning leadership to the seller or buyer shifts transaction prices and payoffs in that agent's favor. Controlled ablations examine how response credit, policy consistency, and reward timing affect learning.