Last week I mentioned the likelihood that 2024 would see the introduction of humanoid robots from at least 2 American companies. These new robots would be able to independently control and coordinate their sensors and actuators in such a way as to perform useful work without direct coding or programming. On Sunday, January 7th, Brett Adcock, CEO of Figure, @adcock_brett, announced that the company (founded in 2022) has successfully merged a humanoid robot with an so-called end-to-end-internal-neural-network-autonomous-control-system. His posts from Sunday on X show a video of the company’s Figure 01 prototype robot making coffee. The accomplishment is not in the making of coffee. The accomplishment is that the robot “learned” how to make coffee and correct its mistakes as it went along, simply by “watching” people do it over a 10-hour period. Adcock did not specify what he meant by “learning” and “watching” since these terms apply to human cognition, not to machines.

Figure-01 has learned to make coffee ☕️

Our AI learned this after watching humans make coffee

This is end-to-end AI: our neural networks are taking video in, trajectories out

Join us to train our robot fleet: https://t.co/egQy3iz3Kypic.twitter.com/Y0ksEoHZsW

— Brett Adcock (@adcock_brett) January 7, 2024

In a second post, Adcock explains:

“Why is this so important? The reason why this is so groundbreaking is if you can get human data for an application (making coffee, folding laundry, warehouse work, etc.) you can then train an AI system end-to-end on Figure 01, there is a path to scale to every use case and when the fleet expands, further data is collected from the robot fleet, re-trained, and the robot achieves even better performance.”

Of course, their prototype is not quite ready to go into production, but production is not far off. Tesla (yes, the car company) is believed to have similar capabilities embodied in their Optimus 2 robot. Pilot production of Optimus 2 robots and their initial application on the Tesla assembly lines is expected this year.

A Note of Caution

This breakthrough has not been independently confirmed. And it is important to clear up some misunderstandings about so-called neural networks. The term “neural” implies a connection to the concept of human neurons. Actually, a neural network would be closer to a vast array of Rube Goldberg machines.

Credit: Rube Goldberg, 1931 Public Domain

Think of each Goldberg machine as determining a single decision in an array of outputs which ultimately make a decision on a much larger scale. In our sample here, this machine will decide whether or not to operate the automatic napkin. We can make the desired outcome more or less likely by adjusting every step in the process. For example, if we change the weight of the seeds which drop into the pail, we can change the likelihood that the napkin will move. If we add more Goldberg machines to our neural network, we can model every factor that goes into your decision whether to finish your meal and go to bed, or to leave the table and drive to Denver. Since some factors will be more important than others (for example whether or not you are sleepy) we can give more or less weight to each factor. If we itemize, model, and adjust the weight of each and all of the factors that go into the decision, and tweak every model appropriately, we can appear to mimic the human decision-making process to a very high degree of accuracy, as long as there is real correspondence between the real world and the synthetic world of the Rube Goldberg machine.

However, the machine is not making a human decision. The entire network of Goldberg machines—the so-called neural network—is not conscious, is not human, but is really just a very complex system of switches that appear to model the outcome of a decision-making process, but not the decision-making process itself. It is a probabilistic process—not a conscious process, and certainly not a human creative process. It is a system of pattern recognition which results in a “monkey see, monkey do” machine control system. That sounds primitive. It is. However, controlling robotic machinery in this way is a huge step forward for control systems which had been limited to simple repetition of either recorded or programmed actions.

For a good technical (more accurate and less humorous) review of neural networks, see the IBM page on neural networks here.

Implications

If the report from Brett Adcock is true, this portends a huge change in industrial production. In the early days of the electrification of America, the occupation of electrician became the biggest category of employment in the nation. Don’t be surprised if the humanoid robot industry becomes the biggest industry in the world by many measures.

The impact of humanoid robots will also upshift human employment categories, and vastly lift productivity and manufacturing, while lifting up the American standard of living. More and more, workers will be responsible less for repetitive work on product, and more and more responsible for overseeing and tending to their robot platoons and making improvements to product and process.

As the age of humanoid robots develops, assuming that the new Trump administration is on top of the situation (remember that the British Imperial predators are always looking for an angle by which to sabotage progress), the incentive will be to build manufacturing facilities inside the United States, so that industry can take advantage of the cultural advantages of the American workforce. The pull of “cheap labor” will be ended as American labor productivity will go through the roof.