Survey candidate RL environment suites not yet covered by EnvPool

[Trinkle23897]

Goal

Keep a lightweight survey of RL environment suites that are popular or increasingly visible in the research community, but are not already advertised as supported by EnvPool.

This is a discovery / prioritization issue, not a commitment to implement every suite below.

Scope note

I intentionally left out the families that EnvPool already advertises in the README / registration entrypoint: Atari / ALE, MuJoCo Gymnasium tasks, Gymnasium-Robotics, classic control, toy text, ViZDoom single-player, DeepMind Control Suite, Box2D, Procgen, and MiniGrid / BabyAI.

Portability triage legend

🟢 source-port-ok means the upstream project appears to use an Apache-2.0 / MIT / BSD-style source license that can be incorporated into an Apache-2.0 project, provided we keep upstream copyright, license, attribution, and per-file notices.

🟡 needs-asset-or-simulator-audit means the project source is permissive, but a working environment may require separately licensed assets, ROMs, scene datasets, maps, fonts, robot models, game binaries, proprietary simulators, or GPU-only backends. Do not vendor those extras into EnvPool until they are audited individually.

🔴 do-not-port-source means we should not translate or vendor upstream source into EnvPool. Prefer a wrapper, a clean-room reimplementation of the benchmark idea, or no support.

Robot learning / manipulation / embodied AI

Related open EnvPool issue: #178 asks for guidance/custom support for complicated robotic manipulation environments.

• Isaac Lab — unified robot-learning framework on NVIDIA Isaac Sim; high interest for GPU simulation, locomotion, manipulation, and sim-to-real. Repo: https://github.com/isaac-sim/IsaacLab
  • 🟡 Needs asset/simulator audit: Isaac Lab framework is BSD-3-Clause and the isaaclab_mimic extension is Apache-2.0, but Isaac Lab requires Isaac Sim; its README points to additional/proprietary terms for Isaac Sim / Omniverse Kit / 3D models / textures. cuRobo has an additional restrictive NVIDIA license. Do not vendor Isaac Sim, cuRobo, NVIDIA assets, or binaries into EnvPool.
• ManiSkill — GPU-parallel robotics simulator + manipulation benchmark on SAPIEN; popular for manipulation and imitation/RL evaluations. Repo: https://github.com/haosulab/ManiSkill
  • 🟡 Needs asset/simulator audit: Code is Apache-2.0 and README says rigid-body environments are under fully permissive licenses, but assets are CC BY-NC 4.0. Port task/state logic only, or require user-supplied/downloaded assets after an asset audit.
• MJLab — MuJoCo-Warp robot-learning environments with an Isaac Lab-style API; a separate task implementation from MuJoCo Playground / MJX and Isaac Lab. Repo: https://github.com/mujocolab/mjlab
  • rl_games usage: Go1 / G1 velocity tracking and Lift-Cube-Yam manipulation, with published training and evaluation results: https://github.com/Denys88/rl_games/blob/master/docs/MJLAB.md
  • 🟡 Needs asset/simulator audit: Apache-2.0 source. Preserve third-party/per-file notices and audit robot XMLs, meshes, textures, and dependencies separately. Decide whether a native CPU EnvPool port is useful relative to the MuJoCo-Warp execution path.
• WujiHand / wuji-mjlab — independent dexterous-hand tasks built on MJLab; keep separate from MJLab's built-in tasks, MyoSuite, and MuJoCo Playground. Repo: https://github.com/wuji-technology/wuji-mjlab
  • rl_games usage: WujiHand_Reorient in-hand cube reorientation, with reported training comparisons and sim2sim evaluation: https://github.com/Denys88/rl_games/blob/master/docs/MJLAB.md#wujihand-in-hand-cube-reorientation
  • 🟡 Needs asset/model audit: Apache-2.0 source. Audit the Wuji Hand model XMLs/meshes and third-party assets individually, and retain their notices; a permissive root source license does not establish redistribution rights for every asset.
• MuJoCo Playground / MJX tasks — DeepMind's newer GPU-accelerated robot-learning environments; should be considered instead of treating legacy Brax envs as the primary target. Repo: https://github.com/google-deepmind/mujoco_playground
  • 🟡 Needs asset/simulator audit: Repository content is Apache-2.0; README says the rough-terrain texture is CC0. Some locomotion/manipulation environments download MuJoCo Menagerie assets on first load, so each model/asset needs a separate license check before vendoring. Also decide whether a CPU EnvPool port is meaningful versus MJX/Warp execution.
• robosuite — modular MuJoCo-based robot manipulation benchmark; classic manipulation baseline suite. Repo: https://github.com/ARISE-Initiative/robosuite
  • 🟢 Source-port-ok with attribution: License file is MIT, and notes partial DeepMind MuJoCo / Apache-2.0 implementation. Keep MIT + Apache notices on any translated files. Audit robot/object model assets separately if copied.
• Meta-World — multi-task / meta-RL manipulation benchmark; overlaps with MuJoCo at the physics layer but is a separate task suite. Repo: https://github.com/Farama-Foundation/Metaworld
  • 🟢 Source-port-ok: MIT source. Keep notice; still audit any copied MuJoCo XML assets/textures separately.
• MyoSuite — musculoskeletal motor-control environments and tasks simulated with MuJoCo. Repo: https://github.com/MyoHub/myosuite
  • Related open EnvPool issue: [Feature Request] Myosuite support #315
  • 🟢 Source-port-ok with asset/model audit: Apache-2.0 source. It is MuJoCo-based; audit anatomical model XMLs, meshes, textures, and any downloaded assets if copying them rather than recreating task definitions.
• RLBench — large-scale robot task benchmark based on CoppeliaSim / PyRep. Repo: https://github.com/stepjam/RLBench
  • 🔴 Do-not-port-source: RLBench uses a custom Imperial College license that is non-commercial/internal-or-academic-only, non-transferable, and non-sublicensable. Do not translate or vendor RLBench source into EnvPool. CoppeliaSim/PyRep/assets would also need separate review.
• Habitat-Lab — embodied AI / navigation / rearrangement task framework. Repo: https://github.com/facebookresearch/habitat-lab
  • 🟡 Needs asset/simulator audit: Habitat-Lab code is MIT, but it depends on Habitat-Sim and common task datasets/derived models are tied to scene-dataset terms; README calls out Matterport3D/Gibson ToS and CC BY-NC-SA 3.0 for those task datasets/models. Do not vendor scene datasets or trained models.

Multi-agent / social / strategy / sports

• PettingZoo environment suites — common multi-agent API + bundled multi-agent benchmark families; EnvPool has multi-player infrastructure but not the PettingZoo suite as such. Repo: https://github.com/Farama-Foundation/PettingZoo
  • Related open EnvPool issue: [Feature Request] Support PettingZoo for Multi-agent testbed #96
  • 🟢 Source-port-ok at framework level; audit per env family: License file says Farama-owned content is released under MIT and some code is inherited under MIT/Apache-2.0. Keep upstream notices and audit each bundled env's assets.
• dm_control locomotion / Soccer — multi-agent continuous-control football with BoxHead, Ant, and Humanoid walkers; outside the Control Suite task set already supported by EnvPool, and distinct from Google Research Football. Upstream: https://github.com/google-deepmind/dm_control/tree/main/dm_control/locomotion/soccer
  • Existing experimental implementation: Denys88/envpool#1, used by rl_games 2v2 self-play. This is missing from upstream sail-sg/envpool main, not an implementation to start from scratch. The fork documents a fixed pitch and a different Humanoid reset-pose distribution; review those oracle gaps before treating the port as aligned.
  • 🟡 Needs asset/model audit: Apache-2.0 upstream source. Preserve notices and audit walker/model assets and any motion-capture data separately if used.
• DeepMind Lab2D — customizable 2D platform for agent-based AI research; it is the substrate/runtime under Melting Pot. Repo: https://github.com/google-deepmind/lab2d
  • Related open EnvPool issue: [Feature Request] Support for DM Lab2d Environments #308
  • 🟢 Source-port-ok with third-party/runtime audit: Apache-2.0 repository. Audit bundled third_party components and decide whether EnvPool should port Lab2D's substrate/runtime or expose a batched wrapper.
• Melting Pot — DeepMind suite for multi-agent / social RL test scenarios. Repo: https://github.com/google-deepmind/meltingpot
  • Related open EnvPool issue: [Feature Request] Melting Pot Integration #163
  • 🟢 Source-port-ok with attribution: Apache-2.0. Audit its substrate/runtime dependencies before vendoring anything beyond task logic.
• Google Research Football — football / soccer RL benchmark with single-agent and multi-agent modes. Repo: https://github.com/google-research/football
  • Related open EnvPool issue: [Feature Request] Google Research Football Integration #151
  • 🟡 Needs asset/simulator audit: Root repo is Apache-2.0; bundled football engine license file is Unlicense; bundled fonts are SIL OFL. If porting renderer/data/assets, keep those licenses and audit any other game data. Source logic appears feasible to port with attribution.
• SMACv2 / SMAC — StarCraft Multi-Agent Challenge family; classic MARL benchmark, but check maintenance and StarCraft dependency cost before prioritizing. Repos: https://github.com/oxwhirl/smacv2 and https://github.com/oxwhirl/smac
  • Related open EnvPool issue: [Feature Request] SMAC Integration #156
  • 🟡 Needs asset/simulator audit; likely wrapper-only: SMAC/SMACv2 wrapper code is MIT, but the environment relies on the StarCraft II game, Blizzard/SC2 APIs, PySC2, and downloaded maps. Do not vendor StarCraft II. A clean-room RTS-like implementation is a different environment.
• Gym-MicroRTS — lightweight RTS-inspired RL environment. Repo: https://github.com/Farama-Foundation/MicroRTS-Py
  • 🟡 Needs asset/simulator audit: Python wrapper is MIT. It wraps the separate Java MicroRTS engine via submodule; audit the engine/license before translating or vendoring engine code.
• JaxMARL — JAX-native MARL environment collection and baselines; see whether any constituent envs are worth CPU EnvPool ports rather than a wrapper. Repo: https://github.com/FLAIROx/JaxMARL
  • 🟢 Source-port-ok for JaxMARL-owned source: Apache-2.0. Its SMAX environment is a simplified/vectorized StarCraft-like environment that avoids the StarCraft II engine; audit any third-party env ports individually.

JAX-native / accelerator-oriented suites

• Jumanji — diverse suite of JAX RL environments, including combinatorial / routing / packing-style tasks. Repo: https://github.com/instadeepai/jumanji
  • 🟢 Source-port-ok: Apache-2.0. Porting pure game/combinatorial task logic is license-compatible with attribution; benchmark/dataset downloads, if used, should be checked separately.
• Gymnax — Gym-like JAX environment collection. Repo: https://github.com/RobertTLange/gymnax
  • 🟢 Source-port-ok with per-env audit: Apache-2.0. Some environments are reimplementations/ports of other suites; when copying a specific env, trace that env's original attribution.
• Pgx — vectorized JAX board-game environments for self-play. Repo: https://github.com/sotetsuk/pgx
  • 🟢 Source-port-ok: Apache-2.0. Board-game/task logic can be ported with attribution.
• XLand-MiniGrid — JAX-accelerated meta-RL environments inspired by XLand and MiniGrid; separate from standard MiniGrid support. Repo: https://github.com/dunnolab/xland-minigrid
  • 🟢 Source-port-ok; 🟡 pre-sampled benchmarks need audit: Apache-2.0 source. README says pre-sampled benchmarks are hosted/downloaded separately; audit before redistributing.
• Craftax — JAX implementation of Crafter / NetHack-style open-ended tasks. Repo: https://github.com/MichaelTMatthews/Craftax
  • 🟢 Source-port-ok: MIT. README references an offline dataset hosted separately; do not redistribute it without separate review.

Safety / autonomous driving / constrained RL

• Safety-Gymnasium — safe-RL benchmark with cost signals; successor-style ecosystem around Safety Gym. Repo: https://github.com/PKU-Alignment/safety-gymnasium
  • 🟢 Source-port-ok: Apache-2.0 and built on modern MuJoCo. Audit any large resource packages / vision assets separately.
• highway-env — autonomous-driving decision-making environments. Repo: https://github.com/Farama-Foundation/HighwayEnv
  • 🟢 Source-port-ok: MIT source; mostly Python task logic. Keep notice.
• MO-Gymnasium — multi-objective RL benchmark suite; specialized, but useful if EnvPool wants first-class vectorized multi-reward support. Repo: https://github.com/Farama-Foundation/MO-Gymnasium
  • 🟢 Source-port-ok with per-family dependency audit: MIT source; README notes optional dependency extras for different env families.

Hard exploration / open-ended games

• Minecraft / Project Malmo — Minecraft-based AI experimentation platform; related to Minecraft/MineRL-style foundation-agent environments. Repo: https://github.com/microsoft/malmo
  • Related open EnvPool issue: [Feature Request] Minecraft Integration #186
  • 🟡 Needs asset/simulator audit; likely wrapper-only: Project Malmo source is MIT, but working missions run against the Java Minecraft client plus the Malmo mod and often require desktop/OpenGL. Do not vendor Minecraft binaries, assets, accounts, or proprietary game data into EnvPool.
• NetHack Learning Environment — NetHack benchmark; long-horizon, partial-observation, sparse-reward. Repo: https://github.com/NetHack-LE/nle
  • 🔴 Do-not-port-source: Repository license is the NetHack General Public License; derivatives including NetHack code must follow its terms. Do not translate/vendor NLE or NetHack engine code into Apache-2.0 EnvPool.
• MiniHack — task-generation / sandbox layer on top of NetHack. Repo: https://github.com/facebookresearch/minihack
  • 🟡 Needs asset/simulator audit: MiniHack Python layer is Apache-2.0, but a working native MiniHack environment goes through NLE/NetHack, which is Red above. Port only MiniHack-owned task-generation ideas after separating them from NetHack engine code.
• Crafter — survival / crafting benchmark; consider Craftax above for a newer JAX path. Repo: https://github.com/danijar/crafter
  • 🟢 Source-port-ok: MIT source. Keep notice.

Retro / broad game collections beyond current coverage

• Stable-Retro — maintained Farama fork for retro console games beyond Atari. Repo: https://github.com/Farama-Foundation/stable-retro
  • 🔴 Do not port bundled emulator cores/ROMs; 🟡 wrapper/integration metadata needs audit: Root project/API is MIT, but LICENSES.md lists emulator cores under GPLv2, MPLv2, LGPL, non-commercial Genesis Plus GX terms, non-commercial Snes9x terms, etc. README says most ROMs are not included and users must provide them; the included Airstriker test ROM is non-commercial. Do not translate/vendor emulator cores or ROMs into EnvPool.

Suggested next pass

For each candidate we should record:

• Does EnvPool want a native C++ port, a batched wrapper, or no support?
• License and asset / ROM / simulator-download requirements.
• Observation / action / reward / cost / multi-agent API shape.
• Determinism and reset semantics.
• Rendering requirements and whether headless RGB is feasible on Linux/macOS/Windows.
• Expected bottleneck: physics, emulator, Python task logic, simulator RPC, GPU-only backend, etc.
• A smallest representative task for correctness + benchmark smoke tests.

[liyc-ai]

Really appreciate your work! Could you please consider supporting Meta-World? Thanks again.

[Trinkle23897]

@liyc-ai Thanks. Meta-World is already included in the survey list above. I marked it as a 🟢 source-port-ok candidate (MIT-licensed source), with a follow-up audit needed for any copied MuJoCo XML / texture assets.

[Trinkle23897]

Thanks again for the suggestion. Native Meta-World support has now been merged in #396.

The new implementation runs the MetaWorld v3 task set through EnvPool native C++ runtime and registers tasks as MetaWorld/<TaskName>-v3, for example MetaWorld/DoorLock-v3. The PR also added deterministic rollout tests, official-oracle alignment tests, render coverage, and docs/render comparison output.

It is available from main now and will be included in the next EnvPool release.