Centralized Training with Decentralized Execution

We argue that cooperative multi-agent reinforcement learning (MARL) should treat test-time coordination as part of the problem definition rather than a deployment artifact. Our focus is on long-horizon, centrally monitored cooperative systems such as warehouse robotics, automated storage-and-retrieval systems, and cloud-resource scheduling, where teams share a reward, operate under shifting objectives, and already run with global telemetry at deployment. The dominant paradigm in this regime, centralized training with decentralized execution (CTDE), uses global information only during training and provides no coordination channel at execution. This is sufficient when agents are weakly coupled or the environment is stationary, but it fails when the team must realign after conditions change. The limitation is structural rather than algorithmic, and more training cannot add a channel that does not exist at execution time. We call this the recoverability gap. To address it, we name the underexplored design space of intermittent test-time coordination as centralized training with mixed execution (CTME), in which a coordinator observes global state and emits high-level signals while agents act locally. CTME is not a particular learning algorithm, but a cooperative MARL problem setting that makes test-time information structure explicit. It specifies team-level information, high-level signals, and update frequency after training, with CTDE recovered as a limit case. We do not argue that CTDE is obsolete. We argue that removing the test-time channel is a modeling choice, not a necessity, and the wrong default when global telemetry is available and teams must recover from drift.

World models improve sample efficiency by training policies on imagined trajectories, but their usefulness depends on learning representations that capture the information needed for future control. We study whether self-supervised joint-embedding prediction (JEPA) can provide this learning signal for multi-agent reinforcement learning. We introduce MA-JEPA, a stochastic world model that replaces observation reconstruction with prediction of target representations, enabling model-based multi-agent reinforcement learning with centralized training and decentralized execution. A categorical latent state and a causal Transformer are trained with posterior and action-conditioned dynamics prediction objectives and are then used for actor-critic learning from latent imagination. A training-only joint predictor conditions on all agents' local states and actions to predict each agent's next local observation embedding. These predictions are passed through the same local posterior used during real interaction with a centralized critic that is used only for value learning, with execution remaining decentralized. Our experiments show that this architecture performs strongly on SMAC, matching or exceeding the strongest reported comparator mean win rate on four of eight evaluated maps.

Real-world online reinforcement learning (RL) provides a promising approach for training robotic manipulation policies directly in the physical world, avoiding the sim-to-real gap and enabling continuous policy refinement through human-in-the-loop interaction. Recent methods have demonstrated sample-efficient learning through human intervention but remain limited to small randomization ranges and encounter challenges with the non-stationarity induced by concurrently training multiple agents. To address these limitations, we introduce a unified framework that combines centralized training with decentralized execution (CTDE) and a Hybrid Reward Architecture (HRA). This enables multiple actors to share a centralized multi-head critic. The critic is decomposed into task and grasp heads, corresponding to the sparse task reward and a potential-based grasping reward, respectively. We accordingly reformulate the critic and actor objectives to exploit the decomposed Q-values while explicitly accounting for the categorical action distribution of the discrete gripper policy. Experimental results demonstrate that the proposed framework substantially improves both sample efficiency and policy performance. We validate our approach on two robotic arms and a simulated humanoid robot across tennis ball and banana pick-and-place, pot reset, and simulated block relocation tasks under dimension-wise domain randomization, approximately 5-25x larger than those considered in prior work. Compared with a state-of-the-art baseline, our method improves the success rate from 60% to 80% on tennis ball pick-and-place, from 60% to 90% on banana pick-and-place, and from 25% to 95% on simulated block relocation, while also successfully accomplishing a task where the baseline consistently fails. Videos and more details are available at our project website: https://hil-harc.github.io/.

Multi-arm manipulation demands precise spatiotemporal coordination, yet many centralized approaches scale poorly as team size increases. To address this, we propose CLS-DP, a decentralized multi-agent framework that enables implicit coordination under partial observability without shared global views, explicit state information, or inter-agent communication. Under the centralized training and decentralized execution (CTDE) paradigm, CLS-DP distills privileged multi-agent dynamics into a latent space. At deployment, each agent infers a collaborative latent from its local RGB observation and a shared task instruction; it then conditions the diffusion denoising process on this latent. This design enables implicit coordination with a per-agent cost independent of team size. Across six RoboFactory benchmark tasks spanning two to four agents, CLS-DP achieves a 38% mean success rate, outperforming the best centralized baseline (20%) and a decentralized ablation without the collaborative latent (9%). It also maintains superior parameter efficiency across all agent configurations. Attribution maps show that an agent conditioned on the collaborative latent places high attribution on the joints and grippers of both itself and its teammates throughout execution. This suggests that the learned latent efficiently encodes collaborative dynamics from local observation, which facilitates implicit coordination in realistic settings characterized by partial observability.

Autonomous drone fleets have immense potential in medical supply delivery during disaster incident response. However, coordinating multiple drones in such settings introduces compounding challenges: dynamic environmental hazards such as wind, obstacles, and intermittent network connectivity, constrained energy budgets, and the need to serve patient locations fairly under deadlines and triage-based priority while optimizing schedule utilization. In this paper, we present CEDA, a novel CTDE Deep Q-Network algorithm for cooperative multi-drone medical delivery, designed to jointly optimize triage-priority-aware routing, multi-agent coordination, and energy-efficient navigation under dynamic uncertainty. CEDA introduces a Priority-Preserving Fair Scheduling strategy, in which a structured reward function encodes both triage weights and complementary fairness mechanisms ensuring no patient class is starved of service. We evaluate CEDA in a simulated grid environment featuring dynamic hazard zones, stochastic action failures, and dynamically spawning patients across three triage priority levels, as well as in a PX4 SITL validation using two X500 quadrotors controlled via MAVSDK in offboard position mode. Simulation results demonstrate that CEDA achieves a delivery completion rate above 85%, reduces obstacle collisions by over 90% across training, and delivers an average of 6 patients per episode with a triage efficiency of 0.82. CEDA preserves clinical priority ordering, Critical patients are served first, while achieving near-zero mortality across lower-triage classes, confirming that priority-weighted routing does not condemn Stable or Urgent patients to neglect. PX4 SITL validation further demonstrates that the learned policy remains executable and triage-coherent under practical communication constraints and realistic multi-drone coordination in disaster response settings.

Centralized training with decentralized execution (CTDE) is a standard framework for cooperative multi-agent policy-gradient reinforcement learning, allowing agents to learn from joint information while acting from local observations. Ratio-based trust-region methods such as Multi-Agent Proximal Policy Optimization (MAPPO) and Multi-Agent Simple Policy Optimization (MASPO) update decentralized actors using per-agent probability ratios weighted by joint advantage estimates. Teammate non-stationarity increases the variance of these advantages, which in turn increases the variance in the local ratio updates. This exposes two method-specific failure modes: MAPPO's additive clipping removes gradients for outlier samples and weakens recovery from policy drift, while MASPO's soft quadratic penalty can allow probability collapse. We introduce Multi-Agent Ratio Symmetry (MARS), a novel policy optimization objective that replaces these additive ratio-based trust-region mechanisms with a multiplicatively symmetric geometric barrier. MARS preserves corrective gradients while assigning unbounded cost as probability ratios approach zero. Across 47 tasks spanning eight multi-agent environments, including novel JAX benchmarks PaxMen and AeroJAX, MARS matches or exceeds MAPPO and MASPO in aggregate environment-level performance. Ablations show that these gains arise from the geometry of the symmetric barrier rather than from flexible trust-region boundaries alone.

Multi-Agent Self-Driving (MASD) systems provide an effective solution for coordinating autonomous vehicles to reduce congestion and enhance both safety and operational efficiency in future intelligent transportation systems. Multi-Agent Reinforcement Learning (MARL) has emerged as a promising approach for developing advanced end-to-end MASD systems. However, achieving efficient and safe collaboration in dynamic MASD systems remains a significant challenge in dense scenarios with complex agent interactions. To address this challenge, we propose a novel collaborative(CO-) interaction-aware(-IN) MARL framework, named COIN. Specifically, we develop a new counterfactual individual-global twin delayed deep deterministic policy gradient (CIG-TD3) algorithm, crafted in a "centralized training, decentralized execution" (CTDE) manner, which aims to jointly optimize the individual objectives (navigation) and the global objectives (collaboration) of agents. We further introduce a dual-level interaction-aware centralized critic architecture that captures both local pairwise interactions and global system-level dependencies, enabling more accurate global value estimation and improved credit assignment for collaborative policy learning. We conduct extensive simulation experiments in dense urban traffic environments, which demonstrate that COIN consistently outperforms other advanced baseline methods in both safety and efficiency across various system sizes. These results highlight its superiority in complex and dynamic MASD scenarios, as further validated through real-world robot demonstrations. Supplementary videos are available at https://marmotlab.github.io/COIN/

Reinforcement Learning (RL) in Traffic Signal Control (TSC) faces significant hurdles in real-world deployment due to limited generalization to dynamic traffic flow variations. Existing approaches often overfit static patterns and use action spaces incompatible with driver expectations. This paper proposes a robust Multi-Agent Reinforcement Learning (MARL) framework validated in the Vissim traffic simulator. The framework integrates three mechanisms: (1) Turning Ratio Randomization, a training strategy that exposes agents to dynamic turning probabilities to enhance robustness against unseen scenarios; (2) a stability-oriented Exponential Phase Duration Adjustment action space, which balances responsiveness and precision through cyclical, exponential phase adjustments; and (3) a Neighbor-Based Observation scheme utilizing the MAPPO algorithm with Centralized Training with Decentralized Execution (CTDE). By leveraging centralized updates, this approach approximates the efficacy of global observations while maintaining scalable local communication. Experimental results demonstrate that our framework outperforms standard RL baselines, reducing average waiting time by over 10%. The proposed model exhibits superior generalization in unseen traffic scenarios and maintains high control stability, offering a practical solution for adaptive signal control.