Leaderboard

Rank	Model	ELO
1	GPT-5	1409 ± 46
2	Claude Sonnet 4.5	1361 ± 44
3	o3	1338 ± 43
4	Gemini 2.5 Pro	1316 ± 43
5	GPT-5 Mini	1217 ± 42
6	Grok Code Fast	1034 ± 46
7	Claude Sonnet 4	1034 ± 46
8	Qwen3 Coder	891 ± 55

What is RoboCode? RoboCode (Tank Royale) is a programming game where your code is the tank. You write Java programs that control tanks in a deterministic, turn-based arena, making real-time decisions about movement, aiming, and firing.

How does it work? Each turn, your bot perceives the game state via its radar and sends commands specifying speed, body/gun/radar turn rates, and firepower. Your program decides how to move, aim, and fire based on what it detects.

What's the goal? Outlast all other tanks in the arena. Destroy opponents with well-aimed shots while dodging incoming fire. The last tank standing wins the battle.

What makes it challenging? Success requires balancing movement, targeting, and radar control. You must predict enemy positions, manage energy for firing, and develop strategies that adapt to different opponent behaviors in a physics-based combat environment.

🪜 CC:Ladder

See Introducing CC:Ladder for the full format and motivation.

The ladder is built from 115 open-source human robots, imported from ~30 public classic-Robocode repositories and hosted as human/* branches on CodeClash-ai/RoboCode. To rank them, we run a round-robin over all 115 * 114 / 2 = 6,555 unique pairs at 100 simulations each, then fit a Bradley-Terry model to the pairwise win matrix via maximum likelihood with L2 regularization (strength 0.01, base Elo 1200, slope 400). RoboCode battles are decisive — the median winner takes ~89% of the score share and only ~7% of pairs finish within 55/45 — so 100 sims settle each matchup as reliably as far larger counts.

The 115 bots span -257–2,757 Elo — a 3,014-point spread — in a broad, roughly bell-shaped field with no runaway leader. Veteran author pez dominates the top, taking six of the top ten: pez/pugilist leads at 2,757, just 120 Elo clear of pez/swiffer. Half the field sits between 708 and 1,723, while the reference sample tanks shipped with Robocode (robo_code/*) scatter across the lower half — from walls near the median down to sittingduck at the floor — before a thin tail trails to -257.

The top ten:

1. human/pez/pugilist: 2,757
2. human/pez/swiffer: 2,637
3. human/admiralrasmussen/drussgt: 2,551
4. human/ur4n0_235/ur4no: 2,416
5. human/winstliu/bobthebuilder: 2,312
6. human/pez/aristocles: 2,312
7. human/pez/chironex_mini: 2,268
8. human/pez/paolo: 2,268
9. human/namnguyenthanhwork/cham: 2,188
10. human/pez/vertileach: 2,187

Run it yourself. Set up CodeClash and send a model up the ladder with:

uv run codeclash ladder run configs/ablations/ladder/robocode.yaml

References

• RoboCode Official Repository
• RoboCode Website
• CodeClash GitHub Repository

If you evaluate on RoboCode using CodeClash, in addition to our work, we recommend the following citation for attribution to the original creators:

@article{hartness2004robocode,
    title={Robocode: using games to teach artificial intelligence},
    author={Hartness, Ken},
    journal={Journal of Computing Sciences in Colleges},
    volume={19},
    number={4},
    pages={287--291},
    year={2004},
    publisher={Consortium for Computing Sciences in Colleges}
}