A subsidiary of Google DeepMind has published a study entitled “Movement in a complex environment”. The paper describes how the method of trial and error helps computer navigate the unfamiliar and challenging spaces. Now this technique is tested in virtual space, but perhaps in the future, it will help robots to walk up or down stairs. Just watch this video:.
All the movements that makes the stickman in this video, he learned by himself. Jumps, bends, jumps — all of these behaviors computer invented independently, calculating the best way to get from point a to point B. Programmers only gave DeepMind AI virtual sensors (that tells him the current position in space) and forced to run forward. Next steps the computer takes himself, coming up with various movements by trial and error.
The novelty lies in the fact that scientists study how complex conditions can teach the AI difficult and deliberate movements. Usually a matter of trial and error causes the subjects to act carefully, resulting in an unfamiliar environment, they are lost, as children who are able to climb the house stairs, but do not understand how to use the escalator. The study showed that this method does not always lead to such results. On the contrary, he can teach the subject of complex movements.
Materials on the subject:.
Video: Jumpy, the robot opens the door and jumps the stairs.
Video: a swarm of small robots are better than one big.
Video: Boston Dynamics “walking” robot in the woods.
What is computer vision?.