The robot hand problem is the hardest problem

Everyone’s watching robots walk. I’m watching their hands.

Walking is mostly solved. It’s hard engineering, don’t get me wrong. But bipedal locomotion on flat and moderately uneven surfaces is something a dozen companies can now do reliably. The algorithms exist. The actuators exist. The balance controllers exist. Walking is a checkbox that most serious humanoid robot companies have checked.

Hands are a different story.

The anatomy of the problem

A human hand has 27 degrees of freedom. 34 muscles control those degrees of freedom, half of them in the forearm (connected to the fingers by tendons, like a marionette controlled from a distance). The fingertips have about 2,500 nerve endings per square centimeter. The hand can apply enough force to crush a soda can and enough precision to thread a needle.

You adjust grip pressure about 80 times per second. You don’t know you’re doing this. When you pick up a glass of water, your hand is constantly microadjusting force based on the weight, the slipperiness of the glass, whether the glass is wet, whether it’s full or empty. All of this happens below conscious awareness. Your hand is running a real-time physics simulation that would crash most computers.

Now build that in metal, plastic, and sensors.

Where the industry is

Figure AI has a hand that can grasp rigid objects. Tesla Optimus can pick up and place objects but the manipulation is still coarse. Shadow Robot Company has the most dexterous commercial hand available, with 20 degrees of freedom, but it costs as much as a car and isn’t ready for industrial deployment.

OpenAI’s Dactyl project showed that a robot hand could learn to manipulate a Rubik’s cube through reinforcement learning. That was 2019. Five years later, no commercial robot can manipulate deformable objects (fabric, food, paper) with anything close to human reliability.

The gap between what a research hand can do in a lab and what a production hand can do in a factory is vast. Lab demonstrations use motion capture, controlled lighting, known objects, and infinite time. Factories have clutter, variable objects, dust, grease, and a requirement to move fast.

Why hands matter more than legs

A robot that can walk but can’t manipulate objects is a very expensive pedestrian. A robot that can manipulate objects but can’t walk is a very capable factory arm (and we already have those).

The value proposition of humanoid robots is that they can do what humans do, in environments built for humans, using tools built for humans. That means opening doors, pressing buttons, turning valves, picking up boxes, folding clothes, cooking food, handling tools.

All of these require hands. Not just grippers. Hands. Multi-fingered, force-sensitive, dexterous hands that can switch between a power grip and a precision pinch in milliseconds.

MIT CSAIL published a paper recently on tactile sensing that gives robots something like a sense of touch. Pressure-sensitive skin that can detect an object’s texture, temperature, and whether it’s about to slip. The results are promising. But going from a research paper to a production-ready hand that works for 10,000 hours in a warehouse is an engineering journey of years.

The company that solves hands

This is my prediction for the humanoid robot industry: the company that solves the hand problem wins the market.

Not the company with the best walking. Not the company with the best language model integration. Not even the company with the best price point. The company whose robot can reliably pick up, manipulate, and place arbitrary objects in unstructured environments.

Because once you have reliable hands, every human task becomes accessible. Warehouse work. Kitchen work. Construction. Agriculture. Elder care. The list of things a robot can do is directly proportional to the dexterity of its hands.

I’ve been tracking progress. It’s slow but real. The reinforcement learning approaches are getting better. The tactile sensors are getting cheaper. The actuators are getting more precise. I’d estimate we’re 3-5 years from a commercially viable robot hand that can handle most rigid objects and maybe 7-10 years from one that can handle fabric and food reliably.

Those timelines are guesses. I might be too optimistic. I’m often too optimistic about robots.

But every time I watch a Figure or Tesla demo, I don’t look at the legs. I look at the hands. That’s where the real race is being run.

Related thinking: