I've been working on a little open-source experiment called llms-robot-arena.
The idea is pretty simple:
Give different AI coding models the exact same rules, ask each one to write the brain of a robot, and then let the robots fight.
The AI is not controlling the robot live. It writes the code first, then the match runs entirely on that code.
All robots are physically identical, so the only real difference is the strategy the model came up with.
They have to figure out things like when to attack, when to retreat, how to conserve energy, how to use recharge zones, avoid holes and dangerous floor tiles, and deal with an arena that becomes more hostile as the match goes on.
What I find interesting is that different models can end up producing noticeably different behaviors even though they all receive the same rules.
Some play aggressively, some are cautious, some seem surprisingly good at positioning, and some make completely terrible decisions.
The matches are reproducible, so the same bots and the same seed always produce the same result.
There’s a live version here:
https://nigrosimone.github.io/llms-robot-arena/
And the project is open source:
https://github.com/nigrosimone/llms-robot-arena
I’m still experimenting with the rules, but the main goal is to make AI comparisons a bit more fun to watch than another benchmark table.
Curious what people think: would you consider something like this an interesting way to compare coding/reasoning models?
Top comments (0)