Optimus walked on stage. Watch its hands
A year after the dancer in the suit, Tesla showed a real humanoid prototype. It walked slowly and waved. The legs get the attention. The hands are the harder and more important problem.
At Tesla’s AI Day on Friday, an actual Optimus prototype walked onto the stage, untethered, slowly, and waved at the audience. Tesla also showed a second version, closer to what they plan to build, which had to be carried out and couldn’t walk yet. Musk said the goal is to make millions of them eventually and sell them for under $20,000.
Last year I wrote that the dancer in the spandex suit was a joke but the idea was right. The prototype is still very early. Boston Dynamics’ Atlas does parkour, and Optimus shuffles. But going from a costume to a walking prototype in a year is fast, and Tesla’s argument that it can reuse its actuators, batteries, vision system and manufacturing experience is at least plausible.
The legs got most of the commentary. I think walking is the easier half. Locomotion is well understood in comparison. There are good controllers, reinforcement learning in simulation works well for it, as DeepMind’s agents showed in 2017, and legged robots from several companies already walk on rough terrain. It’s hard engineering, but people know what the path looks like.
The hands are different. Tesla said the production design has hands with 11 degrees of freedom driven by 6 actuators, with tendons, so some fingers move together. They showed the hand design and some video of it holding objects. A human hand has around 27 bones and more than 20 degrees of freedom, dense touch sensing, and a nervous system that spends a huge share of its motor cortex on controlling it.
Why hands matter so much: almost every useful thing a humanoid would do in a factory or a home involves manipulation, not walking. Picking up objects of different shapes and weights, using tools, opening containers, handling cables and cloth. Grippers can do a narrow slice of that. Human-like hands can in principle do all of it, because the world was designed around them. And in the other direction, a lot of what humans understand about objects, like weight, friction, softness and how things come apart, came from handling them. A robot with crude hands will have a crude understanding of objects.
The two problems with hands are hardware and data. Building a hand that is strong, fast, sensitive and durable, at a low cost, is a very hard mechanical engineering problem. Most research hands break often. Then you need to teach it, and teleoperation data for dexterous hands is slow to collect. Simulation helps, as OpenAI showed with Dactyl, but contact physics is where simulators are weakest.
If I were grading humanoid companies over the next five years, I’d mostly ignore walking demos and look at what their hands can do with objects they’ve never seen.