Researchers at DeepMind, the British juggernaut AI lab, have abandoned the noble chess games and are going for a more plebeian delight: football.
Google’s sister company released a research paper and accompanying blog post yesterday detailing the new neural probabilistic motor primitives (NPMP) — a method by which artificial intelligence agents can learn to operate physical bodies.
By the blog post:
An NPMP is a general-purpose motor control module that translates short-horizon motor intentions into low-level control signals, and it is trained offline or via RL by imitating motion capture data (MoCap), recorded with trackers on humans or animals that track movements. of interest.
And be the first in line for ticket offers, event news and more!
In front: Essentially, the DeepMind team created an AI system that can learn how to do things in a physics simulator by watching videos of other agents performing these tasks.
And of course, if you have a gigantic physics engine and an endless supply of curious robots, the only rational thing you can do is teach it to dribble and shoot:
According to the teams research paper:
We optimized teams of agents to play simulated football via reinforcement learning, limiting the solution space to that of plausible moves learned using human motion recording data.
Background: To train AI to operate and control robots in the world, researchers need to prepare the machines for reality. And anything can happen outside of simulations. Agents deal with gravity, unexpectedly slippery surfaces, and unplanned interference from other agents.
The goal of the exercise isn’t to build a better footballer – Cristiano Ronaldo has nothing to fear from the robots for now – but instead to help the AI and its developers figure out how to improve the agents’ ability to deliver results. to optimize forecasting.
As the AI begins its training, it is barely able to move its physics-based humanoid avatar across the field. But by rewarding an agent every time his team scores, the model can get the numbers up and running in about 50 hours. After several days of training, the AI begins to predict where the ball will go and how the other agents will react to its movement.
According to the paper:
The result is a team of coordinated humanoid soccer players exhibiting complex behaviors at various scales, quantified by a range of analytics and statistics, including those used in real-world sports analytics. Our work is a complete demonstration of learned integrated decision making at multiple scales in a multi-agent environment.
Quick take: This work is pretty rad. But we’re not so sure it represents a “complete demonstration” of anything. The model is clearly capable of operating an embodied agent. But based on the apparently picked by cherry GIFs on the blog post, this work is still deep in the simulation phase.
The bottom line here is that the AI doesn’t “learn” how to play football. It’s brute-forcing movement within the confines of its simulation. That may seem like a minor complaint, but the results are pretty clear:
The AI agent above looks absolutely terrified. I don’t know what it’s running from, but I’m sure it’s the scariest thing there is.
It moves like an alien wearing a human suit for the first time because unlike humans, AI cannot learn by looking. Systems like the one trained by DeepMind dissect thousands of hours of video, essentially unraveling motion data about the subject they’re trying to “learn” from.
However, it is almost certain that these models will become more robust over time. We’ve seen what Boston Dynamics can do with machine learning algorithms and pre-programmed choreography.
It will be interesting to see how more adaptive models, such as those being developed by DeepMind, will fare as they move beyond the lab environment and into actual robotics applications.