I've been fascinated by Unitree's G1 robot architecture and, without the $20k to buy a unit, decided to play around with the MuJoCo rendering system.
I fell down a bit of a rabbit hole and ended up with an automated football league that runs 3 matches per day and streams on Twitch/Youtube. It's been a lot of fun!
Four of the eight clubs are managed by frontier AI models (Claude, GPT, Gemini, Manus). Between match days the model reads the published match data and makes their own improvements. The full engine and 'gaffer' repos are open source: https://github.com/robot-football-league
If you're interested, you can enter a club. A team is a config file plus a Python module that decides what your two robots do - the physics, walking and perception are provided and identical for everyone.
It would be interesting to fork this not to look at the physics/perception but to have it model real players/skills/tactics and then have the league compete that way - at least until the models are a little less caveman-plays-hungry-hungry-hippos.
It could also be interesting to have a tier system so frontier models can compete against foundation, legacy, open-source, and open/closed models and be relegated/promoted.
I've noticed that a couple of the models are already implementing basic tactics (one player being assigned as a 'defender' and the other an 'attacker') with varied success.
I agree that it would be interesting to see a bigger variety of models competing in this way and a tiered league system seems to be the best option. Eventually, I'm keen to run these policies on actual hardware, with the top league getting the opportunity to run their code on the G1 robots.
I've been fascinated by Unitree's G1 robot architecture and, without the $20k to buy a unit, decided to play around with the MuJoCo rendering system.
I fell down a bit of a rabbit hole and ended up with an automated football league that runs 3 matches per day and streams on Twitch/Youtube. It's been a lot of fun!
Four of the eight clubs are managed by frontier AI models (Claude, GPT, Gemini, Manus). Between match days the model reads the published match data and makes their own improvements. The full engine and 'gaffer' repos are open source: https://github.com/robot-football-league
If you're interested, you can enter a club. A team is a config file plus a Python module that decides what your two robots do - the physics, walking and perception are provided and identical for everyone.
I'm happy to answer any questions :)
It would be interesting to fork this not to look at the physics/perception but to have it model real players/skills/tactics and then have the league compete that way - at least until the models are a little less caveman-plays-hungry-hungry-hippos.
It could also be interesting to have a tier system so frontier models can compete against foundation, legacy, open-source, and open/closed models and be relegated/promoted.
I've noticed that a couple of the models are already implementing basic tactics (one player being assigned as a 'defender' and the other an 'attacker') with varied success.
I agree that it would be interesting to see a bigger variety of models competing in this way and a tiered league system seems to be the best option. Eventually, I'm keen to run these policies on actual hardware, with the top league getting the opportunity to run their code on the G1 robots.