Reset audit

Content reviewed 8 November 2026. How this is assessed.

Gantovar and this reference are run by the same operator. Gantovar is excluded from this page, and this page ranks nothing.

Results by family

Each number is counted from results.csv. Package versions run: gymnasium 1.4.0, gymnasium-robotics 1.4.2, minigrid 3.2.0, pettingzoo 1.27.0.

Replay status across all 333 rows: identical 255, replay failed 1 (written check_failed in results.csv), not run 77.

The API check returned pass 176, warnings 80 across the 256 environments run. Replay and the API check are separate results, so an environment can pass one and fail the other.

65 environments raised an assertion error when created, on a default install of the listed package versions, which resolved mujoco 4.0.0. All came from gymnasium_robotics/utils/mujoco_utils.py. They are reported as not run, and versions were not pinned to work around this.

classic/hanabi_v5 replayed with an error: replay failed: ValueError: when an agent is dead, the only valid action is None at pettingzoo/utils/env.py:217. Its reset with no seed began a different game than the one played when the actions were saved, so a saved action became illegal, which means the replay could not finish. It is counted among those that did not replay identically.

12 ids need arguments or a dependency this audit does not install, so they were not run.

Replay counts only when all three seeds match in both processes. A blank or failed replay is in the errored or timed out column.

How the replay was run

For the latest registered version of each environment, every phase ran in a fresh process with a 180 second timeout. An environment that exceeds it is recorded as timeout, and no value is estimated.

Replay: for seeds 0, 1 and 12345, 1,000 actions were sampled once and saved, then replayed in two separate processes with PYTHONHASHSEED 1 and 2. Each run called reset with the seed, stepped through the saved actions, and called reset with no seed after any episode ended. Observations were reduced to data type, shape and raw bytes, recursing into dictionary and tuple spaces. Those bytes, the reward as a double and both end flags went into a SHA-256 chain, and the two chains were compared step by step.

Seeded reset compares the observation from a first reset with a seed against the observation from a reset with the same seed after 1,000 steps. Speed ran serially after everything else, on a machine not reserved for this work.

PettingZoo environments use api_test and the agent-by-agent API. Their actions were sampled once with action masks and saved, so a replay feeds the same actions in the same order. Speed for those environments includes mask sampling and counts one agent turn as one step.

The environment checker was disabled on the make call used for replay and speed, and run separately as the API check on the unwrapped environment with the render check skipped. Numeric thread counts were set to one per process.

What this does not show

This is one machine, on one date, with these package versions. A different chip, build or math library may produce different bytes.

Bitwise identity is stricter than statistical reproducibility. An environment that diverges at step 800 can still give the same learning curves on average. Diverged does not mean broken.

The Arcade Learning Environment and the arcade family in PettingZoo are excluded because they depend on separately supplied game ROM files.

No third-party repository was installed. Only the Farama packages listed here were run, and the exact dependency set is in the requirements files below. Metadata on third-party projects, with no code run, is on the third-party environment status page.

Files

results.csv and results.json hold the full output. reset_audit.py is the harness, run-log.txt is the per-phase log, and the installed package sets are core, robotics, minigrid and pettingzoo.

Every environment, one row each

Sorted alphabetically. Each row has its own link. One registered id is shortened in this table to leave out a hardware maker's name, and the full id is in the CSV. Steps per second is the median of 3 runs of 5,000 steps. Reset time is the median of 20, in milliseconds.

Sources for this entry

• Gymnasium env_checker source: The check_env function and its skip_render_check option, read in the installed 1.4.0 package.
• Gymnasium Env API: Reset, step, seeding and action space sampling.
• PettingZoo environment tests: The api_test and parallel_api_test checks.

Continue this reading path

• Best RL Environments category guide
• How to choose: an RL environment provider
• Resource directory

Cite this entry

Best RL Environments. "Reset audit." Content reviewed 8 November 2026. https://bestrlenvironments.com/reset-audit/