Google's DeepMind AI Learns How to Walk

Posted on July 14, 2017

Google's DeepMind AI has been learning how to walk. This video shows it successfully walking a pair of legs, a human-like figure and an ant-like four-legged creature over different simulated terrains. Google says DeepMind learned on its own how to move these figures and navigate them around obstacles through a trial-and-error process without receiving specific instructions.

ExtremeTech reports that the goal of the project was for the AI to get its 3D figures as fast and as far as it could through the parkour-like courses. The AI learned as it progressed and once successful it would repeat maneuvers that had worked in the past.

The simulated human moves its legs in a familiar manner but its arms go all over the place. The AI often has the human flailing its arms randomly in the air and holding its arms over its head which would not be very energy efficient. The video also never shows DeepMind using the humanoid arms to pull itself up over barriers. It is possible the AI never needed to use the human's arms to complete its mission.

The research paper, "Emergence of Locomotion Behaviours in Rich Environments," can be found here. Here's a video of some of the simulations.