On the humanoid form

A lot has been written about the robotics and physical AI ecosystem that is coming, and many people are focused on humanoid robots. Right now there is a lot of money in it. Humanoid-specific startup funding reached roughly $4.3 billion in 2025. A robot shaped like us can use a world designed for humans, so we avoid having to redesign the environment around it. We also wouldn’t need a specialized robot for every task, one robot could handle many chores, a helpful and versatile assistant. That is interesting and reminds me of Bicentennial Man, among other stories.

But the technology needed to make humanoids work is hard. Rodney Brooks has a nice piece on this, Why Today’s Humanoids Won’t Learn Dexterity. In his words, “Human dexterity relies on a rich sense of touch.” He shows two videos from Roland Johansson’s lab at Umeå University. In the first, a person picks a match out of a box and lights it in seven seconds. In the second, her fingertips are anesthetized, and she fails several times to pick up a match and takes four times as long to light one.

Most humanoid companies are not training for touch though. They train on video of people doing tasks, or on teleoperation where the operator gets very little force feedback. Brooks says Figure and Tesla are both betting on video, and that companies are “trying to train robots to do dexterous manipulation by just showing them the motions, and not getting them to use any force or haptic feedback.” This is the data problem from the previous post, only worse. A video shows where the fingers go, but not how hard they press or when the object starts to slip. Speech and images had decades of microphones and cameras recording them before anyone trained a model on them. For touch, Brooks writes, “We as a species have not developed technologies to capture touch, to store touch, to transmit touch over distances and time, nor to replay it to either ourselves or other humans.” So there is no pile of touch data to start from. The robots would have to generate it, with sensors at a level he says no one has built yet, even in the lab.

Dexterity is one part of it. The other is reliability, whether the robot can do useful work shift after shift. At RoboBusiness in October, Brooks put the bar at 99.999% task success and about 30,000 hours between failures, and said a recent Chinese humanoid hand averaged about 200 hours, according to William Ryu’s summary of the panel. Herbert ten Have, in his overview of what 99 people in robotics argue, cites a Fraunhofer IPA benchmark where a Unitree G1 ran for 1 hour 49 minutes while standing and walking. That would require multiple robots to replace a human performing an eight-hour shift.

Ten Have also counts 15,000 to 20,000 humanoids shipped in 2025, but most of them went to labs, training centers and entertainment.

The piece on SoftBank selling its Boston Dynamics stake shows that even Atlas, one of the pioneers of the humanoid race, is not yet doing paid work in factories.

Atlas is not being deployed broadly across Hyundai plants now. It is not replacing teams of workers today. It is not being sold into hundreds of independent factories and warehouses. Hyundai is talking about beginning with carefully selected tasks in its own facilities, years from now.

That is not a criticism of Boston Dynamics. It is simply the reality of industrial automation.

A robot can be technically remarkable and still have no commercial business case.

I agree with this. Factories buy automation for reasons that have little to do with how impressive the robot looks.

The problem with humanoids is not that they cannot walk, lift objects, recognize parts, or perform impressive movements. Atlas has already demonstrated capabilities that would have looked impossible only a few years ago. The problem is that factories and distribution centres do not buy robotics because it is impressive. They buy it because it improves throughput, reduces labour dependency, improves safety, lowers cost, and delivers predictable performance every day.

Given all this, I don’t think we will see fully functional humanoids in the next few years. I assume it will take 10 to 15 years for the technology to mature and even then I don’t expect wide adoption outside confined environments like warehouses, assembly lines and maybe hotels. Amir Savar makes a similar estimate of about 10 years, writing that “The main reason is data.” The argument is similar to what I go through in the previous post.

I expect it to look a lot like the autonomous car industry. Autonomous driving is not perfect today either, but it keeps improving, and adoption is still not widespread. Ten Have includes one old forecast for calibration, “the 2016 forecast of 600,000 autonomous vehicles a year by 2025.” The last part of the work, mostly on safety, is still left to be done. For humanoids the standards are not there yet. No standard assigns a safety level to a learned policy, and as ten Have points out, “a walking biped cannot simply cut power: it falls.” The first safety standard written for humanoids, ISO 25785-1, is not expected before 2028.

Even if the technology matures in 10 to 15 years, I don’t completely believe we will have humanoid robots in the form the market expects.

If you think about it, minus the legs, you end up with two arms in one robot. That is not new either. ABB’s YuMi was quite unique for its time, a single robot with two arms and some of the dexterity we now expect from a humanoid. Baxter, from Brooks’s earlier company Rethink Robotics, had two arms as well, though according to an ASME piece its customers did not know what to do with that and preferred a single arm. Today’s humanoids have at least that upper body, and then legs. I can make several arguments for why legs are not needed in many tasks, especially in assembly, manufacturing and warehouses with large flat floors. Legs help on uneven terrain and through different kinds of openings and doors, but apart from that they don’t really play a role, in my view. Even Agility Robotics, which builds a legged humanoid, is exploring wheeled robots for warehouses, Jonathan Hurst said at the same RoboBusiness panel. Brooks’s article also discusses why they are not necessary. A CSIS ChinaPower report on Chinese robotics makes the point from the market side, it is “not inevitable that commercial logic would converge on humanoid form factors”.

The same SoftBank piece makes this argument in more detail, and I agree with most of it.

A humanoid is essentially trying to become a universal substitute for human physical work. That is a far more difficult goal than automating a defined process.

It is one of the most complex, expensive, and difficult approaches imaginable to solving problems that purpose-built automation already solves better.

The alternative it describes is to redesign the process instead of copying the person.

The industry keeps talking about humanoids because factories, warehouses, and buildings were designed around people. That part is true. But the logical conclusion is not necessarily that we should build robots shaped like people. In many cases, the better answer is to redesign the process so that people do not need to walk, bend, lift, search, carry, or perform repetitive work in the first place.

That is what good automation design has always been about.

The best warehouse automation does not attempt to reproduce a warehouse worker. It removes unnecessary travel, reduces touches, simplifies material flow, automates repetitive decisions, and brings products to the point where work can be completed efficiently.

The best manufacturing automation does not attempt to duplicate a factory operator from head to toe. It isolates the specific motion, task, or handling problem and designs the most practical machine around it.

So I believe we will have different types of robots rather than one humanoid form. Matt Mason made the same case at an IROS debate on generalist and specialist robots this year. His answer was that “some robots should be generalists, but not all”, and he used the kitchen as an example, “you are the generalist, and your dishwasher, rice cooker, blender, etc. are the specialists.”

However, I think the amount of investment going into humanoids, though absurdly large, is a net positive, and the outcome will be tremendous in a different sense. We are going to see technologies come out of it that are really useful in other applications as well.

For example, I have recently seen a lot more work on sensors, touch sensors and force-feedback sensors, and on algorithms for controlling robots based on force. This has been done for a long time. Rethink built Baxter and Sawyer on series elastic actuators that doubled as force sensors, so the robot could feel its way into a fixture or stop when it hit someone. The key selling point of Universal Robots’ UR series was that it was collaborative, sensitive enough to touch that it could work next to people. But the amount of current work and the money flowing into it are really high now. A Baillie Gifford piece on physical AI notes that “[d]epth cameras and tactile sensors are increasingly affordable and yet more capable.” It is also close to the gap Brooks points at, the missing sense of touch. So I expect to see development of many more of these sensors. I’m sure they will find application in different areas. I came across one where force sensors were used in a medical application, on patients’ feet to understand how force is spread when walking for people with medical conditions.

The other thing I expect in the near term is more small robotic systems adapted for small-scale assembly tasks. Universal Robots launched its first UR robot in 2008 and later the UR3 in 2015, which is small and quite desk-friendly. Part of why UR grew was an open architecture that made it easy to add third-party sensors, cameras and grippers, so integrators could fit the arm to a specific area. I remember working on the UR during my master’s thesis. It was easy to program by teaching it, and we had several custom designed tools as well for it. Compared with the larger ABB or KUKA industrial robots at that time, the UR was a breath of fresh air to work with in close proximity. That category has grown since. With more open-source variants coming in, these small robots need not take the standard robot arm form, they can take different forms adapted to specific applications. They will be specialized to their fields, but they will become more common. I see this as an offshoot of the humanoid work because the parts will come from there, the grippers, gripping sensors and actuators being developed for humanoid hands. They will find a home in a variety of applications outside of humanoid robots as well.

So even though I don’t see humanoids deployed in the field soon, I am hopeful about how this work will reshape other domains, much like Willow Garage’s PR2 did about fifteen years ago. ROS was built alongside it and went on to reshape robotics software well beyond that one robot, and gave rise to many robots within major research labs and industries.

I am more curious about the impact on the software and control side. Right now there is a big push towards physical AI and learned techniques, with the promise of general intelligence, where you train on one robot and can then give it different tasks. This sounds great for the humanoid form, one robot that can replace humans at many tasks. But if the form ends up different, we need to rethink what this work means for the software and control side. At the same time we have many years of experience bringing classical control into these robots. Ken Goldberg calls these two cultures. I think bringing the two together, classical control and generative AI methods, is what will get robots into mainstream use.

Back to all posts