The Problems With Humanoid Robots
I was asked recently if we were building a humanoid robot at Collaborative Robotics. No… simply put, I don’t believe in humanoid robots. I don’t really believe in robotic dogs or cats or horses either.
That’s not to say I’m not impressed by the technology behind these initiatives. Spot Mini, a four-legged dog-like robot by Boston Dynamics is a beautiful feat of engineering. It is as impressive in my mind as the iPhone or the FreeStyle Libre Continuous Glucose Monitor, but nowhere near as useful as those.
There are three specific problems with humanoid robots. One first I believe will be overcome with continued advances in AI. The second might be overcome with enough investor dollars. The third is the Achilles heel.
Boston Dynamics Atlas is similarly impressive in just the power of its actuators and the quality of the robotic controls routines Boston Dynamics is able to demonstrate. Agility Robotics Digit is an amazing robot as well and a gorgeous piece of engineering. The Shadow Dextrous Hand is a similarly gorgeous piece of engineering.
These aren’t the only players in the market, but they’re far ahead of everything else out there. There are some great robotics videos out there (even without putting a human in a robot costume), but a great video is not a production-ready solution. When Wired reported on Boston Dynamics Atlas performing a parkour routine, they were the only journalists to dig deeper to learn that demonstration only worked about 1 out of 20 times. Agility has recently demonstrated the ability to stand back up from the ground; impressive and important, but also reinforces a fundamental problem… their robots can fall down.
There are three specific problems with humanoid robots. One first I believe will be overcome with continued advances in AI. The second might be overcome with enough investor dollars. The third is the Achilles heel.
- The AI isn’t there yet. We lack the generalized controls necessary for robust balancing systems to work in production environments.
- Hardware investments when the AI isn’t there are bad investments. The dollars required to bring a humanoid robot to production quality are likely to be well over $1B invested.
- Biomimicry isn’t the right approach. Humanoid robots aren’t the right design solution for most production tasks.
Let’s talk about the first. Robust controllers for robots are hard. I went to a Stanford-sponsored controls colloquium last year. In one of the talks, Dr Stephen Boyd, one of the true luminaries in the space of controls research and engineering gave an interesting talk. In fairness to Dr. Boyd, I’ll summarize my take-aways which may be different than what he hoped to convey. But overall the talk compared a number of controls techniques, including reinforcement learning, and articulated clearly how they could be reduced to problems of convex optimization, greatly simplifying the problem space. But then he said something interesting (paraphrasing, but hoping I got this right), “so just get the dimensionality under 6 and these problems become classically solvable.”
That was a big a-ha for me. This is exactly what we do in robotics. We reduce the dimensionality of the problem down below 6 degrees of control actuation and we derive a controller using some combination of math, convex optimization, RL or equivalent techniques. Quad-copter drones are 4 degrees of actuation and generally an IMU. Cars are throttle, brake, steer. Airplanes are generally aileron, rudder, elevator, throttle. What Agility has done beautifully is simplify the physics of walking such that the controller can be modeled as a spring-mass system. What Boston Dynamics has done, impressively, is demonstrated the ability to transition from one control regime to another seamlessly, but each controller is simplified. Successful hand controllers in production have reduced the dimensionality with eigenhands, or lower-dimensional controls spaces.
We will eventually get more robust robotics controllers, but there’s a reasonable argument that this problem is as hard, or maybe even harder given the open-ended nature of the world, as developing a self-driving car.
Even when we simplify and reduce the dimensionality of our controls space, we’re still solving controls problems one at a time. We don’t have a ChatGPT-like foundation model that can open any door handle. We need RL or other techniques to derive the motions for opening different types of door handles. TRI has recently shown great promise with learning tasks in a faster, more data-efficient way with their Diffusion Policy work, but they’re also still solving problems one at a time, though admittedly faster.
As I said up front, I think advances in ML/AI will address this problem. We will eventually get more robust robotics controllers, but there’s a reasonable argument that this problem is as hard, or maybe even harder given the open-ended nature of the world, as developing a self-driving car. For instance, self-driving cars are passively stable. They don’t care if they’re transporting liquids or solids, mass sloshing doesn’t affect them. But if a humanoid robot is carrying a box with a bowling ball in it, the controls problem just got very very hard. Humans stabilize our bodies with a lot of different muscles, including our neck muscles which subtly refine the position of our head to keep our center of mass above our feet. That’s super hard to do! And look, we still can’t put a timeline on robust AI for self-driving cars.
But the biggest problem is that humanoids are the wrong solution for most tasks.
This brings us to our second problem. Hardware is expensive. And complex hardware is really expensive. Combining complex hardware engineering costs with open-ended, unsolved AI problems, means the funding requirements are open-ended. And it’s not like you can do some work with a humanoid without solving the balancing problem. I suppose some humanoids are just using a wheeled base, but they’re not intrinsically stable… their center of mass is still too high to be safe.
Is there enough money in the venture ecosystem to make a dent in this? Probably, though Softbank has some of the deepest pockets and thrown a lot of money at robots. Google as well. The returns for those investments to date are more than a little disappointing.
But the biggest problem is that humanoids are the wrong solution for most tasks. Not all tasks, I do think Disney’s animatronic actors will become more and more sophisticated and impressive. In Toyko, there’s a hotel where animatronic dinosaurs check you in. An animatronic human might be a little friendlier than a dinosaur. But when it comes to doing real work in the world around us, biomimicry isn’t the answer.
Wheels are the right answer in logistics, in manufacturing, in hospitals, in airports, in stadiums, on the sidewalk, in office complexes, and in nearly every commercial environment.
Let’s take transportation as an example. For nearly 5,000 years, the horse-drawn carriage was state of the art in land-based movement of goods and people. The Romans built over 250,000 miles of roads to ease the movement of supplies. When cars came along, they needed to work with the existing infrastructure. But Henry Ford famously said “If I had asked people what they wanted, they would have said faster horses.” Faster mechanical horses weren’t the right answer for transportation. Wheels were the right answer.
Wheels are the right answer in logistics, in manufacturing, in hospitals, in airports, in stadiums, on the sidewalk, in office complexes, and in nearly every commercial environment. Also, passive stability, having at least 3 points of contact on the ground, preferably 4, is extremely valuable. Keeping the payload inside the cone of stability rather than cantilevered in front of a robot is better as well.
I truly believe that our approach at Amazon was the right one… focus on understanding what robots are capable of today, deploy current state-of-the-art robots at scale quickly, and then leverage the advances in machine learning and AI to further improve and streamline the operations. That’s the pragmatic approach to bringing robots into the world and with over 750,000 robots deployed, Amazon is far and away the most successful company in the world at deploying robots at scale. I remain simultaneously proud and inspired by the work those teams are doing and look back fondly at my time driving forward that acceleration.
At Cobot, we’re developing a collaborative robot to bring those benefits outside of the four walls of an Amazon fulfillment center or sort center or air hub and into the world around us. We’re not trying to solve open-ended AI problems, we’re doing pragmatic things like using wheels, four points of contact on the ground, bringing payloads into the cone of stability. But we’re doing that in a way that is trustworthy, collaborative and able to work in human designed spaces. We’ll be in our first field deployment in January solving real-world problems in the movement of human-scale loads, namely boxes, totes and carts.
So no, we’re not building a humanoid. As bullish as I am on the future of robotics to improve the world around us, I don’t believe humanoids will be the dominant form-factor. We’re not ready to reveal our form-factor yet, but we’re looking forward to that day.
Brad Porter is the CEO and Founder of Collaborative Robotics, Inc, a Sequoia, Khosla and Mayo Clinic backed robotics company headquartered in Santa Clara, California. Prior to founding Cobot, Brad was the Vice President and Distinguished Engineer leading a global team of 10,000 people overseeing all of Amazon’s logistics robotics work. Brad also served as CTO of Scale AI, Platform Architect for Tellme Networks and an early engineer at Netscape. Brad holds a Bachelors and Masters in Computer Science from MIT. His research focused on computer graphics under Professor Seth Teller.