Deep reinforcement learning or Deep Reinforcement Learning is one of the most promising research fields in the field of artificial intelligence. Basically, through this technique, the system learns to optimize a decision process: if the result of that decision is beneficial, the system automatically learns to repeat that action in the future, while if the result is detrimental, it will avoid taking the same action again. action and replace it with a different one.
Through this system is how a group of researchers from Google has made a robot learn to walk on its own , through a long process of successes and errors.
This learning system is the most appropriate in such a situation because learning to walk is a very complex process marked by countless variables. For that reason, researchers have looked at how children learn to do it: through many trials and falls, until they succeed .
Deep RL
One of the great milestones that has been reached with this development is that the robot, when falling, gets up on its own and continues to try to walk , gradually learning to do better. On the contrary, the algorithms developed so far allowed robots to learn to walk, but they needed the help of a human to get up after a fall.
In this way they created an autonomous locomotion system that allows the four-legged robot, thanks to Deep Reinforcement Learning (Deep RL), to learn to walk on its own.
To improve the algorithm, the scientists delineated the terrain that the robot could explore and trained it in multiple maneuvers simultaneously. Thus the robot overturned 33 times, fell 16 times and got up by itself in a test in which the unit, if it reached the edge of the terrain while walking forward, had to change direction and begin to learn to walk backwards. .
Second, the researchers also restricted the robot’s test movements, making it cautious enough to minimize damage from repeated falls. During the times when the robot inevitably fell anyway, they added another hard-coded algorithm to help it roll back.
Through these various adjustments, the robot learned to walk autonomously across several different surfaces, including flat floors, a memory foam mattress, and a cracked doormat. The work shows the potential for future applications that may require robots to navigate difficult and unfamiliar terrain without the presence of a human .
Thus, through this autonomous process, without the assistance of anyone, the robot took only two hours to learn to walk on flat ground . As the researchers who have developed the system write:
“Our method can acquire a stable gear from scratch directly in the real world in about two hours, without relying on any model or simulation, and the resulting policy is robust to moderate variations in the environment.”
By allowing robots to learn more autonomously, they will thus be closer to being able to learn in the real world we live in.
Reference: Learning to Walk in the Real World with Minimal Human Effort, arXiv: 2002.08550 [cs.RO] arxiv.org/abs/2002.08550
