Although robot dogs are highly capable, they are often not comparable to real animals. Part of the reason is that it’s hard to learn how to walk like dogs directly — but the study from Google’s Artificial Intelligence Lab makes it much easier. The Google team partnered with the University of California, Berkeley, to find a way to enable quad-legged dogs to perform “agile behaviors” such as light trots like real dogs. As the researchers note in their blog post, the established training process often “requires a lot of expert insight and often requires a lengthy reward adjustment process for each desired skill.” “
The dog may not be in full agreement with the robot dog’s action mode, resulting in the latter falling, locking, or other aspects of failure. Google’s AI project solves this problem, adding a bit of controllable chaos to the normal order. Typically, the dog’s movements are captured and key points such as feet and joints are carefully tracked. In digital simulation, these points are simulated as a robot dog, and the virtual version of the robot dog will try to imitate the dog’s movements with its own movements, learning to do.
Despite the researchers’ good results, problems can arise when trying to control an actual machine dog with simulated results. The real world is not a two-dimensional plane, there are no idealfriction rules or anything like that. Unfortunately, this means that uncorrected analog gait tends to cause the robot dog to fall directly to the ground.
To prevent this from happening, the researchers introduced random elements to the physical parameters used in the simulation, making the virtual machine dog heavier, either the motor weaker, or more friction with the ground. This makes machine learning models that describe how to walk take into account the various small variations and how to counteract them.
By learning to adapt to this randomness, the learning walking method is more robust in the real world, so that you can imitate the target dog’s walking style, and can even imitate more complex movements, such as turning and rotation, without any human intervention, only a little extra virtual training. Of course, you can add manual adjustments if needed, but as things stand, this is a big step forward from the previous fully automated action.
In another study described in the same article, another group of researchers described a robotic dog that could walk on its own, but was implanted with the ability to avoid walking outside a designated area and climb on its own when it fell. With these basic skills, robotic dogs are able to continuously improve their training areas without human intervention, learning quite impressive motor skills.