An AI learning system will be developed that can learn human movements and learn soccer movements in 3 days
AI development company DeepMind, an affiliated company of Google, has developed AI technology that can learn from ``how to stand up and walk'' to ``how to play soccer'' with reference to human movements. With the developed learning method, it is possible to learn movements closer to actual humans and animals than conventional methods, and the released movie shows AI playing soccer using dribbling and passing. increase.
From motor control to team play in simulated humanoid football | Science Robotics
From motor control to embodied intelligence
https://www.deepmind.com/blog/from-motor-control-to-embodied-intelligence
Watch how an AI system learns to play soccer from scratch
https://techxplore.com/news/2022-09-ai-soccer.html
Having AI learn the movements of humans and animals is of great help when developing real-world robot motion systems. However, the conventional learning method of ``learning a large amount of data and optimizing it to achieve a goal'' often learns strange movements that are difficult to apply to robots in the real world. did. Click on the image below to see an example of strange behavior from traditional learning methods.
In order to solve the problem of `` it is difficult to learn actions that can be applied to the real world '', the research team devised a learning method of `` motion-capturing human actions and making AI imitate those actions. By clicking the image below, you can see the human movement (yellow) and AI (gray) that imitates human movement from motion capture.
In the following movie, you can see how AI learns the behavior of soccer players using the above method and learns from 'how to walk' to 'how to play soccer 2 to 2'.
In the experiment, a human body model with joints in parts such as arms and torso was used.
In addition, the soccer field used for the experiment is surrounded by walls on all sides, and the rules are simplified for the experiment.
Under the above conditions, AI learned actions such as ``stand up'', ``run'', ``dribble'', ``shoot'', ``pass'', and ``block'' from motion capture data.
Before learning starts, the human body model can not even stand up as shown below.
On the third day from the start of learning, basic movements such as 'running', 'dribbling' and 'shooting' became possible.
On the 50th day from the start of learning, it became possible to play soccer close to the real world, aiming for a goal while passing a pass with a teammate.
According to the research team, applying strange movements acquired by conventional learning methods to robots may damage the robots themselves and their surroundings. By adopting the learning method developed this time, it is expected that robots will be able to acquire safe and efficient movements.
Related Posts:
