Reinforcement learning typically relies heavily on a well-designed reward signal, which gets more challenging in cooperative multi-agent reinforcement learning. Alternatively, unsupervised reinforcement learning (URL) has delivered on its promise in the recent past to learn useful skills and explore the environment without external supervised signals. These approaches mainly aimed for the single agent to reach distinguishable states, insufficient for multi-agent systems due to that each agent interacts with not only the environment, but also the other agents. We propose Synergy Pattern Diversifying Oriented Unsupervised Multi-agent Reinforcement Learning (SPD) to learn generic coordination policies for agents with no extrinsic reward. Specifically, we devise the Synergy Pattern Graph (SPG), a graph depicting the relationships of agents at each time step. Furthermore, we propose an episode-wise divergence measurement to approximate the discrepancy of synergy patterns. To overcome the challenge of sparse return, we decompose the discrepancy of synergy patterns to per-time-step pseudo-reward. Empirically, we show the capacity of SPD to acquire meaningful coordination policies, such as maintaining specific formations in Multi-Agent Particle Environment and pass-and-shoot in Google Research Football. Furthermore, we demonstrate that the same instructive pretrained policy's parameters can serve as a good initialization for a series of downstream tasks' policies, achieving higher data efficiency and outperforming state-of-the-art approaches in Google Research Football.

Reinforcement learning typically relies heavily on a well-designed reward signal, which gets more challenging in cooperative multi-agent reinforcement learning. Alternatively, unsupervised reinforcement learning (URL) has delivered on its promise in the recent past to learn useful skills and explore the environment without external supervised signals. These approaches mainly aimed for the single agent to reach distinguishable states, insufficient for multi-agent systems due to that each agent interacts with not only the environment, but also the other agents. We propose Synergy Pattern Diversifying Oriented Unsupervised Multi-agent Reinforcement Learning (SPD) to learn generic coordination policies for agents with no extrinsic reward. Specifically, we devise the Synergy Pattern Graph (SPG), a graph depicting the relationships of agents at each time step. Furthermore, we propose an episode-wise divergence measurement to approximate the discrepancy of synergy patterns.

We present Adaptive Skill Coordination (ASC) -- an approach for accomplishing long-horizon tasks like mobile pick-and-place (i.e., navigating to an object, picking it, navigating to another location, and placing it). ASC consists of three components -- (1) a library of basic visuomotor skills (navigation, pick, place), (2) a skill coordination policy that chooses which skill to use when, and (3) a corrective policy that adapts pre-trained skills in out-of-distribution states. All components of ASC rely only on onboard visual and proprioceptive sensing, without requiring detailed maps with obstacle layouts or precise object locations, easing real-world deployment. We train ASC in simulated indoor environments, and deploy it zero-shot (without any real-world experience or fine-tuning) on the Boston Dynamics Spot robot in eight novel real-world environments (one apartment, one lab, two microkitchens, two lounges, one office space, one outdoor courtyard). In rigorous quantitative comparisons in two environments, ASC achieves near-perfect performance (59/60 episodes, or 98%), while sequentially executing skills succeeds in only 44/60 (73%) episodes. Extensive perturbation experiments show that ASC is robust to hand-off errors, changes in the environment layout, dynamic obstacles (e.g., people), and unexpected disturbances. Supplementary videos at adaptiveskillcoordination.github.io.

Decentralized execution is one core demand in cooperative multi-agent reinforcement learning (MARL). Recently, most popular MARL algorithms have adopted decentralized policies to enable decentralized execution and use gradient descent as their optimizer. However, there is hardly any theoretical analysis of these algorithms taking the optimization method into consideration, and we find that various popular MARL algorithms with decentralized policies are suboptimal in toy tasks when gradient descent is chosen as their optimization method. In this paper, we theoretically analyze two common classes of algorithms with decentralized policies -- multi-agent policy gradient methods and value-decomposition methods to prove their suboptimality when gradient descent is used. In addition, we propose the Transformation And Distillation (TAD) framework, which reformulates a multi-agent MDP as a special single-agent MDP with a sequential structure and enables decentralized execution by distilling the learned policy on the derived ``single-agent" MDP. This approach uses a two-stage learning paradigm to address the optimization problem in cooperative MARL, maintaining its performance guarantee. Empirically, we implement TAD-PPO based on PPO, which can theoretically perform optimal policy learning in the finite multi-agent MDPs and shows significant outperformance on a large set of cooperative multi-agent tasks.

Coordination is a desirable feature in multi-agent systems, allowing the execution of tasks that would be impossible by individual agents. We study coordination by a team of strategic agents choosing to undertake one of the multiple tasks. We adopt a stochastic framework where the agents decide between two distinct tasks whose difficulty is randomly distributed and partially observed. We show that a Nash equilibrium with a simple and intuitive linear structure exists for diffuse prior distributions on the task difficulties. Additionally, we show that the best response of any agent to an affine strategy profile can be nonlinear when the prior distribution is not diffuse. Finally, we state an algorithm that allows us to efficiently compute a data-driven Nash equilibrium within the class of affine policies.