| name | roadmap-marl |
| description | Plan multi-agent and simulator-heavy environments using appropriate APIs instead of forcing them into single-agent SB3. |
| argument-hint | [task-id] |
| allowed-tools | Read, Write, Glob, Bash |
| context | repo |
| agent | three-mlagents-roadmap-marl-planner |
Roadmap MARL Skill
Use this skill for multi-agent games, traffic, ecological systems,
self-driving, social simulation, LLM agents, or external simulator bridges.
Interface Decision
Choose the interface before implementation:
- Gymnasium: one learning agent controlling one policy.
- PettingZoo AEC: turn-based multi-agent environments.
- PettingZoo ParallelEnv: simultaneous-action multi-agent environments.
- External bridge: simulator owns stepping and exposes a documented protocol.
- Roadmap visualization: frontend demo exists, but training is not yet a
scientific baseline.
MARL Requirements
Before marking a task trainable, specify:
- agent ids and lifetimes;
- per-agent observation and action spaces;
- simultaneous versus sequential action semantics;
- reward ownership and global/team rewards;
- termination and truncation per agent;
- wrappers, vectorization, and baseline learner;
- evaluation metrics beyond aggregate reward;
- reproducible seeds and scenario generation.
Baseline Guidance
Use PettingZoo-compatible tooling for true MARL. SuperSuit wrappers may be
useful for preprocessing/vectorization. For serious baselines, document whether
the plan uses IPPO, MAPPO, RLlib, CleanRL-style scripts, or another maintained
open implementation.
Do not quietly convert a MARL problem into a single-agent SB3 task. If a
centralized controller is scientifically intentional, name it as such in the
task card, UI, and evaluation notes.
Existing Roadmap Tasks
Treat these as roadmap unless and until their backend environment and baseline
pipeline are implemented:
foodcollectorfishintersectionminecraftsimcityself-driving-car
