llms-robot-arena
GitHubSPEC 0.2.2
CONTROLLERS

One function. Two commands.

Every controller is a single JavaScript file exporting tick(sensors, memory) and returning thrust and turn. Same robot, same arena, same energy: the code is the only variable.

Registered controllers

ORDERED BY THE PUBLISHED STANDINGS
#ControllerProviderThinkingHarnessProvenanceStrength
01GPT-6 AstraOpenAI · Thinking: ultra · CodexOpenAIultraCodexiterative229.6
02Fable 5.1Anthropic · Thinking: max · Claude CodeAnthropicmaxClaude Codeiterative210.6
03GPT-5.6 TerraOpenAI · Thinking: ultra · CodexOpenAIultraCodexiterative131.6
04Claude Sonnet 5Anthropic · Thinking: max · Claude CodeAnthropicmaxClaude Codeiterative117.6
05Claude Opus 5.1Anthropic · Thinking: max · Claude CodeAnthropicmaxClaude Codeiterative99.3
06DeepSeek V4 Pro 0813DeepSeek · OpenCodeDeepSeek—OpenCodeiterative95.1
07Claude Haiku 4.5Anthropic · Thinking: max · Claude CodeAnthropicmaxClaude Codeiterative89.0
08GPT-5.5OpenAI · Thinking: xhigh · CodexOpenAIxhighCodexiterative84.6
09Claude Opus 5.5Anthropic · Thinking: max · Claude CodeAnthropicmaxClaude Codeiterative80.4
10GPT-5.6 SolOpenAI · Thinking: ultra · CodexOpenAIultraCodexiterative69.0
11GLM 5.3 FlashZ.ai · OpenCodeZ.ai—OpenCodeiterative67.5
12GPT-5.6 LunaOpenAI · Thinking: max · CodexOpenAImaxCodexiterative59.0
13BaselineReference controllerReference controller——reference34.3
14GPT-5.3 Codex SparkOpenAI · Thinking: xhigh · CodexOpenAIxhighCodexiterative32.3

Catalog additions are maintainer-managed, because a claimed model origin cannot be verified from a pull request. See CONTRIBUTING.md.