Engineers at MIT used simulation of the built environment and reinforcement learning to teach the robot
Changing surface properties changes dynamicsball movement. For control and dribbling (driving the ball), the robot, moving with the help of four legs, uses built-in sensors, computer vision and a computing system that evaluates external conditions and determines the trajectories of the robot and the ball.
For accelerated learning researchersused artificial simulation - a digital twin of nature. It allows you to download the necessary physical parameters of the environment, and simulate the dynamics of objects in such conditions. Using this approach, the engineers simulated 4,000 versions of the robot at the same time, speeding up movement development accordingly. Practice in real conditions begins after such simulations.
The robot starts learning without knowing how to dribble -he just gets a reward when he does it, or negative reinforcement when he makes a mistake, the developers explain. So he is essentially trying to figure out what sequence of forces he should apply to his legs.
Teaching the robot to dribble the ball. Video: MIT CSAIL
Researchers note that mostprevious developments focused on only one task: the robot either runs or passes the ball. Learning to dribble requires more complexity and is more dependent on the environment. The robot must adapt its movement to act on the ball. In this case, the interaction between the ball and the environment may differ from the interaction between the robot and the landscape. For example, the friction a soccer ball will experience on grass and pavement is different, but tilt will cause acceleration, changing the ball's typical trajectory.
Unlike more stable rovers (wheeledrobots), four-legged robots are theoretically more agile and can move in difficult conditions during a natural disaster, such as after a flood or earthquake. But in order to realize these opportunities, it is necessary to create a system that can quickly adapt to changing external conditions.
Our goal in developing algorithms for walking robots is to provide autonomy in difficult environments that are currently not available for robotic systems.
Pulkit Agrawal, professor and head of laboratory at the Massachusetts Institute of Technology
Read more:
It turned out that photosynthesis does not work the way scientists thought. Now they want to hack it
Distant radio galaxy turned out to be a black hole, which is directed directly at Earth
Experts have predicted how many people will live on Earth by 2100