MARLHospital: Fair Multi-Agent Healthcare Benchmark

• MARLHospital is a healthcare-inspired multi-agent environment that models fair coordination by integrating agent skill variations, explicit fatigue constraints, and order-dependent tasks.
• It supports diverse team configurations and realistic clinical task sequences, enabling detailed analysis of skill-task matching and energy-constrained scheduling.
• Its composite fairness objective balances workload and skill alignment to prevent burnout and ensure sustained, equitable performance in emergency scenarios.

MARLHospital is a customizable, healthcare-inspired multi-agent environment introduced to study fair collaboration and scheduling in heterogeneous healthcare worker teams, particularly in scenarios resembling emergency departments. It was proposed in the context of FairSkillMARL, where fairness is defined not only as workload balance but also as skill-task alignment, with explicit attention to burnout, overuse of highly skilled agents, fatigue, and structured coordination in team-based clinical procedures. The environment is positioned as a response to the claim that existing simulators are not well-suited to this problem because they do not jointly model varying expertise, order-dependent medical tasks, and shared-task execution under energy constraints (Ekpo et al., 26 Aug 2025).

1. Problem framing and intended scope

The central problem addressed by MARLHospital is the mismatch between conventional fairness formulations in multi-agent reinforcement learning and the operational realities of healthcare teams. In the motivating paper, fairness in MARL is described as being often framed as a workload balance problem, which overlooks agent expertise and the structured coordination required in real-world domains. In healthcare, this is operationally important because equalizing the number of assigned subtasks can still produce task-skill mismatches, excessive reliance on the most capable agents, and downstream burnout risk (Ekpo et al., 26 Aug 2025).

MARLHospital therefore targets heterogeneous healthcare worker (HCW) teams rather than interchangeable agents. Its design emphasizes three coupled properties: per-agent skill variation, fatigue-sensitive execution over long horizons, and task structures in which subtasks may be sequential, shared, or order-dependent. The paper explicitly motivates the environment with emergency-department-like code situations, where correct timing, coordinated alternation, and expertise-aware assignment are integral to performance rather than secondary considerations.

A plausible implication is that MARLHospital is not merely a task-completion benchmark. It is a fairness-sensitive coordination environment in which allocation quality depends on who performs a subtask, when that subtask is performed, and whether prior exertion constrains future actions.

2. Environment design: team structure, task structure, and implementation

MARLHospital supports multiple team composition paradigms so that collaboration can be studied under varying degrees of heterogeneity and dependence. Agents can be assigned individual skill levels, and these skill levels are task-specific rather than purely global.

Team type	Capability pattern	Coordination consequence
Uniform Teams	All agents have identical capabilities	Minimal expertise asymmetry
Specialized Teams	Each agent has clear expertise in at least one subtask but can still perform others less efficiently	Allocation must trade off redundancy and specialization
Interdependent Teams (Forced Cooperation)	Agents can perform only subsets of tasks	Some tasks cannot be performed alone

This structure is used to reconstruct clinical team hierarchies that include both specialization and redundancy. In the specialized setting, agents retain limited cross-task competence, which allows inefficiency rather than impossibility to become the dominant failure mode. In the interdependent setting, coordination is mandatory because task feasibility itself depends on multi-agent complementarity (Ekpo et al., 26 Aug 2025).

The task space is similarly structured. Actions encode both movement and clinically inspired subtasks such as “compress chest,” “give rescue breaths,” and “administer AED.” Tasks are procedurally generated according to real-world codes, including American Red Cross Adult Basic Life Support. The environment distinguishes between shared and individual tasks, and it enforces order dependency: for example, rescue breaths require the CPR sequence first. This makes the environment closer to a workflow domain than to a simple allocation domain, since policy quality depends on respecting temporal and logical dependencies across subtasks.

At the systems level, MARLHospital is built on Robotouille and wrapped for use with the EPyMARL benchmarking framework. This implementation choice gives it direct compatibility with established MARL pipelines while preserving domain-specific task semantics.

3. Energy-constrained scheduling and fatigue modeling

A distinctive feature of MARLHospital is its explicit fatigue model. The paper motivates this component with CPR, where healthcare workers are required to alternate every 2 minutes to prevent fatigue. The environment operationalizes this by assigning each agent a maximum energy, a per-action energy cost, and an energy recovery rate. Shared tasks therefore become scheduling problems under physiological resource constraints rather than purely logical coordination problems (Ekpo et al., 26 Aug 2025).

In practice, chest compressions are modeled as a high-cost shared task. If an agent has sufficient remaining energy, the agent can continue the task; otherwise the agent must rest and recharge. A noop action is available to represent mandatory rest. This creates endogenous turn-taking: alternation is not hard-coded as a symbolic protocol but emerges from the interaction between task demands and agent energy states.

This design changes the meaning of fairness. Without energy constraints, a policy may repeatedly exploit the most skilled agent. With energy constraints, overuse has immediate operational consequences because it reduces future availability and can block completion of later subtasks. The paper links this explicitly to healthcare concerns about fatigue, errors, and burnout. As a result, MARLHospital can evaluate whether a policy distributes effort in a way that remains operationally feasible over long horizons rather than merely optimal at the current timestep.

The experimental setup includes both unconstrained and constrained energy regimes. The unconstrained setting uses energy cost $0$, while a constrained example uses cost $3$ and recharge $1$. This permits direct comparison between coordination problems dominated by sequencing and those additionally shaped by fatigue-sensitive alternation.

4. Formalization and fairness objective

MARLHospital is formalized as a decentralized partially observable Markov decision process (Dec-POMDP),

$(I, S, A, T, R, \Omega, O, \gamma),$

where $I = \{1, \ldots, n\}$ denotes the agents, $S$ the states, $A = A_1 \times \cdots \times A_n$ the joint action space, $T$ the transition function, $R$ the reward, $\Omega$ the observation space, $3$0 the observation function, and $3$1 the discount factor. The state includes 2D spatial location, skill levels, and ongoing actions, while each agent receives only local observations, making coordination genuinely decentralized rather than centrally scripted (Ekpo et al., 26 Aug 2025).

The fairness formulation attached to the environment is the core conceptual contribution. Let $3$2 be the number of agents, $3$3 the number of subtasks, $3$4 the number of subtasks assigned to agent $3$5, and $3$6 the skill of agent $3$7 on subtask $3$8. Workload imbalance is measured by a Gini Index $3$9 over the workloads $1$0. Skill-task misalignment is defined as

$1$1

The composite disparity is then

$1$2

with $1$3 controlling the balance between workload balance and skill alignment. The shaped reward becomes

$1$4

This formulation directly encodes a common misconception that the paper seeks to reject: equal workload is not, by itself, sufficient for fairness in healthcare collaboration. A policy can have low workload disparity while still assigning critical tasks to poorly matched agents or concentrating high-stakes work on a few experts. In MARLHospital, fairness is therefore a dual objective rather than a single balance criterion.

5. Experimental use and reported findings

The environment is used to compare FairSkillMARL in conjunction with four standard MARL methods—IQL, MAPPO, VDN, and QMIX—and against two state-of-the-art fairness metrics. The reported experiments use 3 agents, multiple seeds, and separate training and testing phases, with hyperparameters following EPyMARL best practices. Evaluation tasks include CPR, described as short-horizon and order-dependent, and rescue breaths, described as longer-horizon with more complex sequencing (Ekpo et al., 26 Aug 2025).

Performance is assessed not only by task completion but also by fairness-sensitive diagnostics. The reported metrics include success rate, workload Gini Index, Agent Contribution Dispersion, and Skill-Task Alignment Penalties. This metric suite is significant because it allows a distinction between policies that merely solve the task and policies that solve it while preserving equitable and expertise-aware collaboration.

Several findings are emphasized. Centralized training with decentralized execution performs better than fully decentralized learning, with VDN highlighted as particularly effective when task complexity or energy constraints increase. Forced-cooperation teams are more difficult to coordinate than uniform teams, reflecting the cost of hard capability constraints. The paper also reports that energy constraints can sometimes help coordination because enforced alternation introduces structure into shared-task execution.

Most importantly, the results are presented as evidence that workload-only fairness is insufficient. The paper states that fairness based solely on equal workload might lead to task-skill mismatches, and it reports that incorporating skill alignment yields statistically significant improvements in coordinated team performance and agent equity, especially at a tuned balance of $1$5. This suggests that the environment is useful not only for benchmarking MARL algorithms but also for stress-testing fairness definitions themselves.

6. Position within hospital multi-agent research

MARLHospital occupies a specific niche within hospital-oriented multi-agent research. Other recent work has addressed hospital logistics and operations through distinct formulations: stochastic AMR scheduling with battery, capacity, and time-window constraints solved via variable neighborhood search (Cheng et al., 2023); UAV fleet coordination for time-critical medical supply delivery under a POMDP with PPO-based learning (Guven et al., 11 Mar 2026); township-level epidemic control balancing hospital capacity strain and human mobility through H2-MARL (Luo et al., 13 Mar 2025); and intraday operating-room scheduling under uncertainty using a cooperative Markov game with centralized training and decentralized execution (Liu et al., 4 Dec 2025).

A separate line of work focuses on hospital administration and cross-institutional information systems rather than HCW task allocation. H-AdminSim provides a FHIR-integrated multi-agent simulator for administrative workflows such as intake, scheduling, rescheduling, and cancellation (Lee et al., 5 Feb 2026). Other healthcare multi-agent studies target medical imaging, federated learning, or cross-hospital diagnosis rather than embodied team coordination, including a prior-knowledge-guided deep reinforcement learning framework for thoracic disease classification (Nie et al., 2023), FedMRL for heterogeneous federated medical imaging (Sahoo et al., 2024), and MedLatentDx for privacy-aware latent communication in cross-hospital rare-disease diagnosis (Wang et al., 11 Jun 2026).

Against this background, MARLHospital is best understood as a benchmark and research instrument for fairness-aware team collaboration. Its distinguishing combination of heterogeneous expertise, order-dependent medical tasks, explicit fatigue, shared-task alternation, and composite fairness penalties makes it a vehicle for studying how multi-agent systems allocate effort in settings where safety, sustainability, and expertise alignment are inseparable. This suggests a broader methodological role: MARLHospital can serve as a controlled environment for comparing fairness definitions, coordination mechanisms, and scheduling strategies before such ideas are transferred to higher-stakes clinical decision-support settings (Ekpo et al., 26 Aug 2025).