LK

What a control loop actually is, the five different systems hiding inside one robot, why some robots feel gentle on contact and others don't, and the safety layer that has to think before it stops a robot.

Google DeepMind announced one robotics release this week. Read past the headline and it's four separate businesses wearing one name: a model driving an entire humanoid body for the first time, a planner coordinating more than one robot at once, an offline version that learns a new robot body in hours, and a safety test that checks the other three before they act. Each of those sits in a different slice of the technology stack, and none of them belongs to just one industry, which is exactly the skill this field rewards: knowing which slice, which industry. The full article's in the reply below this post, it maps out exactly how companies like Komatsu, Figure, and two dozen others actually divide, layer by layer, domain by domain. Try it on a company you know: what's holding it back, capital, certification, or supply chain? Reply with your guess. The actual argument, worked through properly with the numbers, is in the reply below this post.

A driverless mining truck has hauled billions of tons of rock for years, quietly profitable. A humanoid robot that costs far more still can't say what it costs to finish one task. Here's one reason why. Get a single action right 95% of the time, which sounds pretty good. String a hundred of them together, the kind of sequence in something as basic as unloading a dishwasher, reach, grasp, lift, place, again and again, and the odds of the whole job going right fall to about 1 in 170. To actually pull it off nine times out of ten, every single action needs to succeed something like 99.9% of the time. Almost nothing in robotics is measured against that number today. That's one of three costs that show up the moment a machine has to act in the real world instead of on a screen. Wrote up the other two, with the numbers and the sources, in the article below.