Introduction
After being away for a few years, I made my way again today to the Brabanthallen in ’s Hertogenbosch, where for the 22nd time this year the Vision, Robotics & Motion fair was organized. Over 120 exhibitors, the vast majority showcasing proven technologies in the field of optimized production and logistics environments. My attention was especially drawn to the premiere at this fair: humanoid robots, i.e., robots with a human form.
Comparing the evolution of humanoid robots with AI
Right at the first keynote by Randall van Poelvoorde , tech expert and visionary of RobotXperience , a comparison was made between the rise of humanoid robots and the exponential growth artificial intelligence has gone through in recent years. The statements during this presentation stuck with me. You cannot see artificial intelligence separately from robots. Whether they have a human form or are built for a specific use case. Why do I think you cannot compare the growth of humanoid robots with that of artificial intelligence (at least in recent years, since neural networks and machine learning have existed since the 1950s)?
Let’s start with the evolution of the human body. Why has our body evolved so much over the centuries? Our body is designed to move functionally. Movement is essential to survival: in searching for food, hunting, fleeing danger, and performing daily tasks. By moving functionally, our body is better prepared for the challenges of daily life, like lifting, bending, carrying, and walking.
The functional design features of the human body form a blueprint for developing humanoid robots. Here, we immediately encounter technical limitations in mimicking the biomechanics of a human body:
Human intuition is difficult to replicate, even using vast amounts of training data and numerous simulations. This goes too deep for this blog post, but research is already underway on how AI can be made more human by combining intuitive models with newly acquired information. Interesting material at the crossroads of neuroscience, developmental psychology, and intuitive physics.
Robots, regardless of whether they have a human form, are not efficient in managing energy. Mammalian muscles still outperform robots in energy conversion. As a result, robots need heavy batteries. For a current comparison: Boston Dynamics’ Spot robot dog consumes about 400W and has an action radius of up to 90 minutes on lithium-ion batteries. A sled dog, however, has about 68 kWh of energy (stored in fat) and can work for days.
Earlier this year, a humanoid robot ran a half marathon in Beijing. Along the way, the battery had to be replaced 3 times for a full charge. On this front, various research directions exist: ranging from robots converting metal into energy to fluid-based energy storage and circulation systems that can increase energy density.
As described, the human body excels at multitasking. This is also one of the reasons, according to Randall, why humanoid robots are built: they can perform dozens of tasks per robot. One moment standing at an assembly line, the next moment cleaning toilets. However, looking at the current state of AI and robotics, there is only one conclusion: the architecture and AI models used are very good at performing one specific task, usually in a controlled and predictable environment.
Although modern robots are becoming increasingly versatile thanks to software updates and simulations, coordinating multiple tasks remains technically challenging. And of course, developments here are ongoing, such as integrating hierarchical control systems and motion programming. Robots are also highly dependent on predefined scenarios. While humans can intuitively (yes, there it is again) switch between tasks and quickly respond to unexpected situations, robots must be explicitly trained or programmed for each task and combination of tasks.
There simply is no AGI (artificial general intelligence) yet that can generalize tasks enough to understand and learn every intellectual task a human can perform without human intervention.
But it’s not only these technical limitations that make me think it will still be a while before humanoid robots are deployed en masse. I also think the following business-related aspects will matter:
Although it seems that a business case for the use of such a robot can be easily made, I still have some question marks regarding the business case for the original manufacturers. Let’s look at some numbers: Tesla, one of the manufacturers of humanoid robots, currently makes about a 7% margin on sold cars. Analysts and investors expect Tesla to achieve higher margins from the mass rollout of robo-taxis. In the most optimistic scenario, 30%, comparable to platform companies like Uber—but then without human drivers and with lower operating costs. So, I also expect Tesla to face a similar challenge with the deployment of their humanoid robots.
It’s not so much about building these robots but more about finding a scalable business model like the platform model of robo-taxis.
Not the underlying AI models themselves, but the user interface and user communities have ensured the exponential growth of ChatGPT and the like. The basis of large language models is a token generator: they predict the most logical next token based on an underlying vector database and statistical models. Users worldwide have built custom GPTs for specific tasks, such as content marketing, data analysis, productivity, research, learning new skills, and hobbies, thanks to the open user interface.
In my expectation, it will be more difficult to build user communities and marketplaces around humanoid robots. There are already closed marketplaces and communities for AVGs, AMRs, and drones. I certainly would not call these successful. Simply because specific technical expertise is required that is not available on every street corner. So, I expect a strong dependency on the original manufacturers to make the robots suitable for customer-specific use cases. Which, again, stands in contrast to finding a scalable business model.
In which part of the world will these robots be deployed first? China has been working on these developments for a decade. Therefore, I expect China—still leading as a production country—to be the first to deploy such robots. Perhaps with less concern about safety standards, ethical aspects regarding labor market impact, and acceptance in the workplace. In that sense, I think our Dutch or even European market will not be early adopters. I consider the recent introduction of humanoid robots, such as at Mercedes-Benz in Berlin, a positive exception.
Eventually, humanoid robots will certainly find their way onto the production floor. I expect a growth pattern similar to what smartphones have gone through in terms of time duration. The supply chain for humanoid robots will become better and more efficient. The underlying technology will make evolutionary steps. But the whole thing just takes time.
Given the current geopolitical tensions and the position Europe plays in it, I am thinking more in terms of decades than a magnitude of a few years. Randall also referred to the vacation economy or the backyard paradise following a question from the audience. It may well take some more time before people massively relax once robots take over our work and work so efficiently that supermarket prices return to 1970s levels. It seems like a very boring society in that way.
And that’s it for today. If you have questions following this blog post, feel free to connect with me.