Google Is Letting Real Dogs Teach Their Robotic Brethren How to Walk

Your puppy may have been a bit clumsy at first, but it probably wasn’t long before little Rover was running around with remarkable agility. A typical doggo can easily scamper over extremely rough terrain without a second thought. Robots, on the other hand, tend to fall flat on their mechanical faces the moment they come across a small pebble. Roboticists have made really substantial progress towards improving that in recent years, but there are still challenges to overcome. To help, a team from Google’s AI lab and the University of California Berkeley are letting real dogs teach their robotic brethren how to walk.

The concept here is almost obvious in retrospect. We have the technology to capture and quantify the minutia of a dog’s movement as it runs. If we create a robot version of that dog, then the robot should be able replicate the real dog’s movements in order to run just as well. This isn’t a new idea, but it hasn’t been particularly successful in the past. That’s because Rover is making tiny adjustments on the fly to adapt his gait to the terrain. He’s doing that using a variety of sensory information, including the feel of the ground under his paws and the terrain he sees in front of him. That sensory information is given relevance thanks to Rover’s previous experience. In contrast, the robot would simply be reproducing whatever Rover’s gait was at the time it was recorded.

To address that shortcoming, the team used simulations based on the real dog’s gait that had an element of randomness introduced. It’s difficult to build a simulation that truly reflects the real world, so they didn’t bother trying. Instead, they simply allowed some variation in the many physical parameters that define the simulation. For example, they might make the virtual robot dog weigh more or less than the real dog in some simulations. Over the course of many simulated training sessions, those parameters were continuously tweaked. This resulted in a much more robust model, which the robot dog was able to take advantage of in order move about with far fewer failures.