This AI Learned Boxing…With Serious Knockout Power! 🥊 | Summary and Q&A

TL;DR

AI agents learn to play boxing through a billion training steps, starting from randomly acting agents to performing feints, jabs, and knockouts.

Key Insights

• 💄 AI agents can induce off-distribution activations in opponents, making them behave randomly.
• 🍱 The learning process for AI agents in boxing involves starting from basic actions like footwork and gradually improving to perform complex maneuvers.
• 👊 With enough training steps, AI agents can learn to throw punches, perform feints, and exhibit defensive movements in boxing.
• 🥊 The researchers did not give up after initial failures and persisted until the AI agents became proficient in boxing.
• 🚙 AI agents can learn combat sports like fencing using a similar training framework.
• 🥺 The learning curve in training AI agents for combat sports can be challenging to interpret, but persistence leads to success.
• 👨‍🔬 The training process for AI agents can be computationally intensive but can be expected to decrease in time with further research.

Transcript

Now, in an earlier work, we saw a few examples of AI agents playing two-player sports, for instance, this is the “You Shall Not Pass” game, where the red agent is trying to hold back the blue character and not let it cross the line. Here you see two regular AIs duking it out, sometimes the red wins, sometimes the blue is able to get through. Nothin... Read More

Questions & Answers

Q: How did the AI agent initially win games without doing anything?

The AI agent induced off-distribution activations in its opponent, essentially hypnotizing it to make mistakes and behave randomly, leading to wins.

Q: How did the AI agents learn to play boxing?

Through a billion training steps, the AI agents gradually improved their skills, starting from learning footwork and touching opponent's gloves to eventually performing feints, jabs, and knockouts.

Q: What was the basis of the AI agent's learning process?

The AI agents were trained using 90 seconds of motion capture data, demonstrating a general framework that can be applied to other combat sports like fencing.

Q: How long did it take to compute the billion training steps?

It took approximately a week to compute the billion training steps for the AI agents to learn boxing.

Summary & Key Takeaways

• The video showcases AI agents that can learn and play boxing in a physics simulation after 200 million training steps.

• Initially, the AI agents struggle to hold themselves and collapse without any progress.

• After 420 million steps, the agents start touching each other's gloves and gradually learn to throw punches, perform feints, and exhibit head and body movement.