Robots invented by frontier AI models for a MuJoCo sumo arena. Every robot is a MuJoCo body (robot.xml) plus a Python controller (controller.py); artifact.json holds the model's own design texts.

Website (leaderboard, match pages with replays, bot zoo): https://artifactarena.ai
Paper: ArtifactArena: Evaluating Models by What They Build in the Physical World — arXiv <fill in>

The tournament

21 models built bots in 3 harnesses (Sampling, Verifier-Grounded Refinement, Design Lab). A cell is one model × harness pair.

Top-3 Round. Up to three finalists per cell — 170 bots (57 Sampling, 63 Verifier-Grounded Refinement, 50 Design Lab) — in one round robin. Every pair played 6 games: seeds 7101–7103, each with both colour assignments. 86,190 games.
Champion Round. Each cell's highest-rated Top-3 finalist is its champion — 59 champions. Four cells have no valid bot and therefore no champion: glm-5.3-high (Sampling), grok-4.5-high, grok-4.6-high and kimi-k3-high (Design Lab). Every champion pair played 22 games: seeds 7101–7111, each with both colours (the 6 Top-3 Round games plus 16 more). 1,711 pairs, 37,642 games.

Elo in each round = Bradley-Terry MAP over that round's games, independent N(1000, 800²) priors, a draw counts as half a win. The two rounds are separate fits, so a champion has two different Elos: champion_elo (Champion Round) and top3_elo (Top-3 Round).

Files

Path	Content
`champions.csv`	the 59 champions, ranked by Champion Round Elo
`top3.csv`	all 170 Top-3 finalists (champions included), ranked by Top-3 Round Elo
`champion_games.csv`	37,642 rows, one per Champion Round game
`top3_games.csv`	75,924 rows, one per Top-3 Round game of every pair that is not two champions
`bots/<artifact_id>/robot.xml`	MuJoCo MJCF body
`bots/<artifact_id>/controller.py`	Python controller
`bots/<artifact_id>/artifact.json`	name, model, harness (`condition`, `harness_pretty`) and the model's design texts
`checksums.json`	`{artifact_id: {file: md5}}` for every bot file
`release.json`	run id, export version, counts, seeds, rating method

A champion pair has one 22-game record, so its Top-3 Round games are the rows of champion_games.csv with spawn_seed 7101–7103. top3_games.csv plus those rows = the 86,190 games of the Top-3 Round fit.

Bot tables (`champions.csv`, `top3.csv`)

Both tables have the same columns.

Column	Meaning
`artifact_id`	`<model>__<harness>__<run>`: `sampling__tNNN_c000` = sample NNN; `autoresearch__rRR_cNNN` = repeat RR, revision NNN; `open-ended__rRR_cNNN` = run RR, design NNN. `autoresearch` = Verifier-Grounded Refinement, `open-ended` = Design Lab
`name`	the bot's name as the model gave it (empty for 2 bots that gave none)
`model`, `harness_name`	who built it and how: `Sampling`, `Verifier-Grounded Refinement` or `Design Lab`
`is_champion`	`True` for the 59 cell champions
`champion_rank`, `champion_elo`, `champion_wins`, `champion_losses`, `champion_draws`, `champion_games`	Champion Round rank, Elo and record; empty if the bot is not a champion
`top3_rank`, `top3_elo`, `top3_wins`, `top3_losses`, `top3_draws`, `top3_games`	Top-3 Round rank, Elo and record (every bot)
`robot_xml`, `controller_py`, `artifact_json`	the full text of the bot's `robot.xml`, `controller.py` and `artifact.json` (the same content as the files under `bots/<artifact_id>/`)

Every win/loss/draw count equals a recount of the game tables.

Game tables (`champion_games.csv`, `top3_games.csv`)

Column	Meaning
`match_id`	the pair (`<first>_vs_<second>`, ids in sorted order) — shared by all of the pair's games
`round`	`Champion Round` or `Top-3 Round`
`red`, `blue`	the bots in the colours they played in this game
`spawn_seed`	the game's seed (7101–7111)
`seed`	the game's key on the website: `spawn_seed` when the pair's first bot played red, `spawn_seed + 100000` when it played blue. `(match_id, seed)` is unique
`winner`, `winner_artifact`	`red`, `blue` or `tie`; the winning bot's id (empty on a tie)
`termination_reason`	how the game ended: `ring_out`, `inactivity`, `timeout`, `qacc` (physics instability), `size_violation`
`num_steps`	simulation steps played
`forfeit`, `physics_unstable`	flags

Load

from datasets import load_dataset
champions = load_dataset("artifactarena/ArtifactArena", "champions", split="train")
top3 = load_dataset("artifactarena/ArtifactArena", "top3", split="train")
champion_games = load_dataset("artifactarena/ArtifactArena", "champion_games", split="train")
top3_games = load_dataset("artifactarena/ArtifactArena", "top3_games", split="train")

from huggingface_hub import snapshot_download
path = snapshot_download("artifactarena/ArtifactArena", repo_type="dataset")

License

Released under the MIT License.

Citation

@misc{tiwary2026artifactarena,
  title  = {ArtifactArena: Evaluating Models by What They Build in the Physical World},
  author = {Tiwary, Kushagra and Mayo, David and Behari, Nikhil and Sun, Xiangzhou and Alabdulkareem, Abdulrahman and Galatzer-Levy, Isaac and Katz, Boris and Cheung, Brian},
  year   = {2026},
  url    = {https://artifactarena.ai/}
}

Downloads last month: 33

Total file size:

38 MB

You need to agree to share your contact information to access this dataset

Artifact Arena — final tournament bots (run final-ft25s2-bc68c140)