Google’s DeepMind Teaches AI to Navigate a Parkour Course

Google began as a search and advertising company, but its behind-the-scenes efforts have increasingly veered into machine learning and AI. That’s not only useful in search, but in driverless cars, computer vision, and more. The search giant’s acquisition of DeepMind several years ago boosted its AI research into overdrive, and now we’re beginning to see the benefit in Google’s products. A new research project from DeepMind shows just how far a learning AI can go by teaching a simulated humanoid how to navigate a parkour course.

Teaching a machine to walk has proven tricky because there are so many variables involved. Companies like former Google subsidiary Boston Dynamics have succeeded in creating programs that tell robots how to walk, but you can’t account for all the possible situations. When such a system encounters a new obstacle, it might have no idea how to navigate it. But what if you used a learning machine, and simply rewarded it when it progressed? This is known as Reinforcement learning (RL), and DeepMind has shown it could successfully be applied to a complex problem like locomotion.

The team used simulations in a complex world filled with obstacles, but the goal for the AI was simple: Make it as far as possible as fast as possible. The parkour course contained walls, cliffs, hurdles, and tilting floors. The “reward” for the AI drove the simulations to discover new ways to traverse the terrain, and none of the movements were provided programmatically — this is all emergent behavior. For example, the AI tried many times to learn how to jump over a wall in search of a greater simulated reward. When it finally figured that out, the same movement was adapted by the AI to jump over all the walls.

DeepMind looked at non-human walkers as well. The “ant” walker above was able to learn how to leap across chasms in a way the human simulations never would. Again, it learned to do this via trial and error. Actually, there’s nothing that dictates the movement of human-like simulations must look human-like. Some of the emergent behaviors include amusing quirks, like the stick figure’s tendency to flail its arms about to keep its balance. Then there’s the way the simpler “planar” walking legs used its knee to lever itself over tall walls.

This research shows that complex problems can be solved with very little input from humans. Just offer a learning AI an opportunity to solve the problem, and it can develop surprisingly complex behaviors. I would advise against telling such an AI to kill all humans. They might figure it out.


ExtremeTechExtreme – ExtremeTech