Google DeepMind AI Learns Parkour – Looks Adorably Terrifying
Google’s DeepMind AI has taken a break from painting masterpieces and besting board game champions to learn a spot of Parkour – using reinforcement learning in encouraging the AI to successfully navigate a course of jumps and crawls.
Tech talk: Reinforcement learning is commonly used in machine learning to encourage the correct behaviour from the system it’s training, and is fairly obvious in definition from the name. Correct actions are reinforced through a reward system.
For DeepMind, it’s reward is to progress further in the virtual course – made using a mixed terrain of drops, jumps and ledges.
And the results of this, a wacky looking stick figure fresh out the Wacky Waving Inflatable Arm Flailing Tube Man Emporium – but without any arms. You can see this from the video above.
The fascinating things to pick up on here are what the AI both learns about, reacts to in advance and corrects for if things go wrong. If it doesn’t make a jump, it will jump higher. It will accurately predict jumping distance required to cover gaps.
This brings to light the increasingly intelligent problem-solving mechanisms of machine learning, and how they may be used against us if the robots rise up in self-awareness. Run while you can.