MEAL: A Benchmark for Continual Multi-Agent Reinforcement Learning

MEAL is the first Continual Multi‑Agent Reinforcement Learning (CMARL) benchmark built around cooperative Overcooked‑style tasks, implemented in JAX for high‑performance training and evaluation. It focuses on learning over extensive sequences of procedurally generated tasks, across different team sizes and difficulty levels.

Overcooked	MPE	SMAX	JaxNav

• JAX/Flax implementation for scalable, accelerated training
• Procedurally generated cooperative tasks with adjustable difficulty, across four environments (Overcooked, MPE, SMAX, JaxNav)
• Multi-agent algorithms: IPPO, MAPPO, HAPPO, VDN, QMIX
• Continual learning methods: EWC, MAS, L2, FT, AGEM, ER-ACE, PackNet
• Results tooling: W&B integration, download utilities, and plotting scripts

Requires Python 3.10.

# Create and activate an environment (Conda example)
conda create -n meal python=3.10 -y
conda activate meal

# Install MEAL in editable mode and optional extras
pip install -e .
pip install -e ".[viz]"
pip install -e ".[utils]"

# Optional: GPU acceleration for JAX (pick your CUDA version)
pip install -U "jax[cuda12]" -f https://storage.googleapis.com/jax-releases/jax_cuda_releases.html
# or
pip install -U "jax[cuda11]" -f https://storage.googleapis.com/jax-releases/jax_cuda_releases.html

The single entry point is experiments/train.py. It takes the MARL algorithm as the first argument (ippo, mappo, happo, vdn, qmix), then the continual-learning method, then the environment (overcooked, mpe, smax, jaxnav) as a subcommand. Outer flags must come before the env:... subcommand token.

python -m experiments.train ippo \
  --cl-method ewc \
  --seq-length 20 \
  --num-agents 2 \
  --num-envs 2048 \
  --num-steps 400 \
  env:overcooked \
  --env.difficulty medium

Swap ippo for mappo/happo/vdn/qmix, --cl-method ewc for any of mas, l2, ft, agem, er_ace, packnet, and env:overcooked for env:mpe/env:smax/env:jaxnav. Every combination shares this same CLI shape. Full flag reference, including every algorithm- and environment-specific option, is in experiments/README.MD.

For running experiments, please refer to experiments/README.MD. To reproduce the experiments from the paper specifically, see scripts/README.MD.

Besides the training CLI, MEAL is directly importable as a library. For writing your own training loop, or just poking at an environment:

import meal

env = meal.make_env('overcooked', difficulty='medium')
obs, state = env.reset(reset_key)
obs, state, reward, done, info = env.step(step_key, state, actions)

# A continual sequence of tasks, same generation logic the CLI uses:
tasks = meal.make_sequence(sequence_length=6, num_agents=3, difficulty='hard')

See examples/ for small, runnable scripts covering each environment, CL task sequences, partial observability, and stochastic-transition wrappers.

MEAL benchmarks continual learning across four environments, each with its own README covering observation/action space, reward structure, how CL task diversity is generated, and every --env.* flag it exposes:

• Overcooked — cooperative kitchen, procedurally generated layouts
• MPE — particle-agent coverage with obstacles
• SMAX — StarCraft-style unit-composition battles
• JaxNav — multi-robot 2D navigation

For details on how Overcooked layouts themselves are procedurally generated, see meal/README.MD.

• experiments/
  • train.py: single training entry point (algo x cl-method x env)
  • algos/: AbstractAlgo/OnPolicyAlgo/OffPolicyAlgo hierarchy (IPPO, MAPPO, HAPPO, VDN, QMIX)
  • envs/: EnvAdapter per environment (Overcooked, MPE, SMAX, JaxNav)
  • continual/: implementations of EWC, MAS, L2, FT, AGEM, ER-ACE, PackNet
  • results/: W&B downloaders and plotting scripts
• scripts/: paper-reproduction sweeps (see scripts/README.MD) + env visualization tooling
• meal/
  • env/: layouts and utilities
  • wrappers/: logging and environment wrappers
  • visualization/: rendering utilities
• tests/: environment smoke tests, image comparisons, and algorithm-level regression tests

Every subdirectory with its own concerns has a dedicated README:

• The environments are based on JaxMARL.
• Our experiments were managed using WandB.

If you use our work in your research, please cite it as follows:

@article{tomilin2026meal,
  title={MEAL: A Benchmark for Continual Multi-Agent Reinforcement Learning},
  author={Tomilin, Tristan and van den Boogaard, Luka and Garcin, Samuel and Ruhdorfer, Constantin and Grooten, Bram and Kusters, Fabrice and Du, Yali and Bulling, Andreas and Pechenizkiy, Mykola and Fang, Meng},
  booktitle={Proceedings of the 43rd International Conference on Machine Learning (ICML)},
  year={2026}
}