💡AI agents can win two-player sports games by doing nothing and inducing off-distribution activations in their opponents.
🥊AI agents can learn to box through joint-actuated characters in a physics simulation.
🥊AI agents start off with little coordination and collapse during early stages of training.
🥊AI agents gradually improve their boxing skills through millions of training steps.
🥊AI agents eventually learn to perform feints, jabs, and knockout punches after extensive training.
How do AI agents induce off-distribution activations in their opponents?
—AI agents induce off-distribution activations by doing nothing in a way that reprograms their opponents to make mistakes and act randomly.
How do AI agents learn to box?
—AI agents learn to box through joint-actuated characters in a physics simulation, gradually improving their skills over millions of training steps.
Do AI agents start off as skilled boxers?
—No, AI agents start with little coordination and often collapse during early stages of training. However, with continued training, they gradually develop boxing techniques.
What techniques do AI agents learn during training?
—AI agents learn various techniques such as feints, jabs, evading punches, and delivering knockout blows.
How long does it take for AI agents to become skilled boxers?
—It takes approximately a billion training steps, which can be computed in about a week, for AI agents to become skilled boxers.
00:09AI agents can win two-player sports games by inducing off-distribution activations in their opponents.
01:23AI agents learn to box through joint-actuated characters in a physics simulation.
02:25AI agents start with little coordination and often collapse during early stages of training.
03:32AI agents gradually improve their boxing skills through millions of training steps.
03:40AI agents eventually learn to perform feints, jabs, and knockout punches after extensive training.
