Of all the jobs we thought AI might steal, boxers weren't high up on the list. But with AI now being used for everything, it was only a matter of time before someone used machine learning to train virtual robot fighters.
Final Automata streams bouts between physically simulated robots that are trained by reinforcement learning. For the moment, the focus is on boxing, but the idea is to train robots in other martial arts and make them fight to see which martial art is supreme.
Final Automata is set in the year 2050. Humans no longer fight each other. Instead they use AI-driven robots as their proxies. It was developed using Unity (see our piece on Unreal Engine vs Unity) by a team of just two people under the name of Chaos Dynamics. It's not a game as you can't control the robots: you just watch them as they fight. Since AI controls both motor movements and strategy and decides every move live, there are no pre-recorded animations and every fight is different.
The only competitors so far in this work in progress are Tuff and Ruff, who, after 250 days of simulation, have mastered basic movements and strategy. The have learned the same moves, but Tuff is more aggressive and Ruff more defensive. In their latest clash, the robots had learned to bounce and were less rigid, now able to rotate nearly 180 degrees. We're told that movement is less random and damage scoring is more realistic than in the first streaming.
It's another interesting example of what AI can do, although I'm not sure the UFC needs to worry too much just yet, but it could be of interest to human coaches and fighters who want to study technique and fighting styles (when the robots get good enough). For now, I'd say that watching AI control simulated characters beat the hell out of each other is likely to remain a niche interest, but I could be wrong. Perhaps we'll see Ruff and Tuff competing in the Olympics. On which note, check out our roundup of Paris 2024 design controversies.
