Researchers at DeepMind, a UK-based artificial intelligence lab, have abandoned the game of chess and focused on the popular sport because of its greater appeal to football. The Google-owned company recently published a research paper along with a blog post about its new Neural Engines (NPMP) that AI can use to learn how the human body physically works.
A part of DeepMind’s blog post states:
An NPMP is a general-purpose motor control module that translates its goals into low-level control signals and is trained offline or through RL by simulating motion capture (MoCap) data obtained by placing a tracker on the human or animal body.
According to the report TNWthe DeepMind team has essentially created a new artificial intelligence that can learn to perform these movements by watching videos of physical body movements and perform them in a physical simulator.
Of course, if you have a giant physics engine and an endless number of curious bots, the only logical thing to do is to teach them how to dribble or shoot.
In a part of the DeepMind team’s research article, it is written:
We optimized teams of agents for a simulated soccer game by reinforcement learning and restricted the solution space to acceptable moves learned using human motion capture data.
It is worth noting that researchers must prepare machines for the real world in order to train artificial intelligence to do work and control robots in the world; This is while in the real world anything can happen and agents have to deal with gravity, unexpected slippery surfaces and unplanned interference from other agents.
The goal of this exercise is not to make a better soccer player; Cristiano Ronaldo Now, he has no fear of bots, and instead videos of his games help AI developers learn how to optimize agents’ ability to predict the best course of action in the real world.
As the AI training process begins, the tool can barely move its physics-based humanoid avatar around the field. After a few days of training, DeepMind’s new artificial intelligence has predicted things like where the ball will go and other factors regarding its movement.
In another part of DeepMind’s article, it is stated:
The result of the development of this artificial intelligence is the creation of a team of soccer players who exhibit complex behavior at various scales that will be measured based on a wide range of analytics and statistics, including real-world sports analytics. Our work is a complete representation of integrated decision making learned at different scales in a multi-agent environment.
We don’t know what DeepMind’s new AI will do once it’s finished. Obviously, this model can work with an embodied agent, but based on the released animations of it, it appears to still be in the simulation stage.
The bottom line is that AI is not learning how to play football. It’s a brutal move at the limits of simulation, and it may seem like a minor crisis, but the results are quite obvious.
The above AI agent looks completely terrified and it is not clear what he is running from. DeepMind’s new AI now moves like an alien wearing human clothes for the first time. The systems that DeepMind trained analyze thousands of hours of video and essentially extract motion data about the position they are trying to learn from.
However, we can say that such models will become more powerful over time. We’ve already seen what Boston Dynamics can do with machine learning algorithms and pre-programmed choreography.
The development of DeepMind’s adaptive artificial intelligence models looks very interesting and it remains to be seen how they will perform in the world of robotics after completing the training course and outside the laboratory environment.