[mac01021]

These press releases from BD are certainly impressive and are always great fun to watch.

But they're mostly silent on the matter of how much autonomy the robots are operating with. I never know how much of their behavior is human directed, or how those directions are conveyed to the robots.

[visarga]

They say that robots are being remote controlled. So, there is an operator telling it to duck, grab, and where to go, but the robot does the actual balancing on itself.

I am a little bit disappointed that they are not using deep learning for robotics. Instead, they are simulating the robot based off a model, using Control Theory. It might be easier to get results from Control Theory but it doesn't offer a path towards more complex behavior, like Deep Learning.

[Animats]

Modern control theory is rather close to machine learning. Adaptive model-based feedforward control is machine learning. The machine learning part builds a model of the dynamics of the system. Then that model is inverted (solved for control inputs) to make it a control system.

They're doing this right. They have a very good basic body control system. Now someone can build higher level strategies to get work done on top of that. That's how biological brains work, after all. Google/Alphabet could, for example, reuse much of their automatic driving software as high level control for this robot.

Google should have BD manufacture a few hundred of those machines, and try to get the cost down to $25K or less per unit for that production run.

[argonaut]

Sure, the brain uses the concept of abstraction, but that is so far away from supporting the assertion that "that's how the brain works."

[Animats]

Mammal brains have multiple functional units. The cerebellum does most of the motor control. The cortex does most of the planning and deciding. The cortex acts through the cerebellum, not by driving muscles directly. Most of Boston Dynamics' control systems are doing cerebellum-level functions. As with the cerebellum, this involves fast control via feedback loops.

[argonaut]

Except for the inconvenient fact that the cerebellum is not explicitly solving control theory equations.

And also that other inconvenient fact that neuroscientists barely understand the brain at all.

[Animats]

"Except for the inconvenient fact that the cerebellum is not explicitly solving control theory equations."

It might be. You can invert a model by training a neural net to compute its inverse.

[mac01021]

Thanks! That seems mostly like what I expected.

I would have to practice with the controls for a really long time to be able to pick up a soda can that deftly, though.

While they are not using deep learning, I think they're doing good work building a platform that can later be controlled via neural net or whatever. The control-theory-based balance and locomotion means that the lowest-level aspects of operation will not demand resources from the neural net and provides the prototypers in the meantime with a way of experimenting to see what body designs might be most effective.

[argonaut]

Deep learning for robotics is tremendously harder than you're making it out to be. It's not magic pixie dust you just sprinkle over something. The state of the art in physical robot deep learning is robots learning a single simple action like picking up something in front of them, and even this is quite difficult.

[jonnycowboy]

BD is only focusing on the platform.

Google is working on the deep learning aspect of finding out what to do / action and how to do it.

The control theory remark above is true, however even with deep learning, you can start your estimate of the parameters with the current control state, which reduces training time significantly.