How We Verified 51 RL Environments (and What We Learned)
Most "awesome RL" lists are link collections nobody maintains. We wanted rlenvlab's directory to be different: every repo link verified live, every paper link checked, maintenance status derived from real commit history. Here's exactly how we did it — and the traps we found along the way.
The pipeline
- Repo existence via the GitHub API. Every one of the 45 unique repos was checked live. No guessing from memory, no trusting old READMEs.
- Redirect resolution. GitHub orgs move. We followed redirects to canonical names:
deepmind/*becamegoogle-deepmind/*, and HighwayEnv and D4RL now live underFarama-Foundation/. - Maintenance from
pushed_at. "Maintained" means a commit within the last 12 months — not vibes. This caught repos that look alive but haven't been touched in years. - Paper links verified against arXiv. Every arXiv ID was checked for existence and title match before linking.
- Star counts as integers. Real current numbers from the API, not rounded figures copied from blog posts.
What we caught
The arXiv check paid for itself immediately: 7 of our 11 uncertain paper IDs turned out to be completely different papers — our guessed "MuJoCo Playground" and "Isaac Lab" IDs pointed at unrelated work. Those entries ship with no paper link rather than a wrong one. We also caught a 404 on what we thought was the Overcooked-AI repo — the real one uses an underscore (overcooked_ai), not a hyphen. These are exactly the errors that accumulate silently in unmaintained lists.
What the data taught us
- 34 of 51 environments are maintained — a healthier ratio than the ecosystem's reputation suggests, thanks largely to Farama's stewardship.
- The graveyard is real but specific: OpenAI Retro, DeepMind Lab, NLE, SMAC, RLCard, and PySC2 are archived or stale. If your tutorial pins one of these, budget time for dependency archaeology.
- Surprises cut both ways: Microsoft's AirSim looked dead but shows commits from September 2026. Reputation is a lagging indicator; commit history isn't.
Keeping it honest
Verification isn't a one-time event — it's a cadence. Our checking scripts live alongside the dataset, and we re-verify on a monthly cycle as part of the changelog. If you spot a dead link or a repo we've misjudged, that's a data bug, and we'd rather hear about it than host it.
Browse the result: the directory, all 51 environments, every link checked.