top of page

Humaniform Robots

December 2023

This piece was originally written and translated into Korean as a part of LG Technology Ventures' Monthly Newsletter to business units and strategic partners. It is republished here with permission. Sensitive information has been removed.

In this article, I refer to human-like robots as “humaniform” rather than “humanoid”, as this was the vocabulary used by Isaac Asimov to create a distinction between human-like robots and human-like aliens.


In most science fiction, there are two views towards robotics - either robots are pervasive, and often human-like in design (humaniform robots) or they are largely absent, with intelligence integrated directly into the items that we use every day (an “Internet of Things” future). After the boom of Internet of Things startups in the mid-2010s, it seemed like the Internet of Things future would be the most likely path, but given the recent fundraising activity and development of humaniform robots like Tesla’s Optimus, Figure, and Agility Robotic’s Digit, it may make sense to revisit that assumption.


The “Internet of Things” Future


Isaac Asimov’s “Robots” series of books envision a future full of robots. In some scenes, robots are so plentiful that there are humaniform robots waiting at futuristic TVs, radios, and even light switches to flick them on and off for humans. To us today this seems a laughable waste of resources, but Asimov explained this development in his book “The Caves of Steel” (around page 170 in most prints):


“The decision was made on the basis of economics. Look here, Mr. Baley, if you were supervising a farm, would you care to build a tractor with a positronic brain, a reaper, a harrow, a milker, an automobile, and so on, each with a positronic brain; or would you rather have ordinary unbrained machinery with a single positronic robot to run them all. I warn you that the second alternative represents only a fiftieth or a hundredth the expense.”

The core assumption in this thesis is that the compute power necessary to automate a device would be extremely expensive compared to the rest of the device. However, in reality, the compute necessary to make a device “smart” is very inexpensive. The cost of parts to make a light switch smart is on the order of a dollar even in small quantities, and even compute necessary for autonomous control of complex machines (such as the tractor and automobile listed as an example above) is less than a thousand dollars. As a result, until now we have seen a future more closely resembling an “Internet of Things” future as inexpensive chips and radios repurposed from the smartphone supply chain are put in every imaginable home device, and high-power GPUs adapted from the gaming industry are used in larger autonomous systems.


The Robotics Future


So why are we excited about humaniform robots today? It turns out that Asimov was almost right. It isn’t compute (or “positronic brains” as he called them) that is expensive - but the actuators that are. For example, laundry machines know that users want to fold laundry after they finish a load but lack the arms to do so. Specialty laundry folding machines proved uneconomical because of the cost and space for the arms and other components necessary to make the product work. At the same time, we have seen the rapid automation of devices like tractors and planes because the actuators (motors, harvesters, sprayers, etc) were already a part of the machine, and all that was needed was a (relatively inexpensive) brain to control them.


The compute necessary to operate robots has been around for decades but has been getting exponentially cheaper, particularly after the the massive improvements in large neural networks and specialty chips to run them. More important than lower cost, the boom in AI capabilities and infrastructure has enabled these robots to learn new tasks, rather than fulfill narrow pre-programmed functions. Figure, for example, is developing models that combine general intelligence with a robot’s own movement models to allow it to understand the context of the world around it and what helpful actions to take. A robot can now understand that a banana peel on the floor is trash and a hazard, and will know that the correct action is to pick it up and throw it away in a trash can. Robotic movements can now be trained through massive parallel physics simulations, such as Nvidia Omniverse, or by analyzing the movements of humans demonstrating a task.


If it is just a question of usefulness, then why haven’t we seen single-purpose humaniform robots? The truth is that we have. Boston Dynamics’ Atlas and Sony’s Asimo (named after Isaac Asimov) are examples of limited-use humaniform robots, but because of their high cost, they never saw widespread adoption, even in narrow pre-programmed tasks. The issue is that the actuators were just too expensive. Efforts to make robotic arms less expensive also had difficulty with market traction because end users generally rely on system integration contractors who are not incentivized to lower costs that they simply pass through to their customers.


Elon Musk notably bemoaned the lack of supply chain necessary to build Tesla’s Optimus, and announced plans to design their own rotational and linear actuator motors optimized specifically for Optimus. Figure is taking a similar approach. Elysium Robotics, C-Motive, Estat, and others are also tackling the problem of robotic actuators through entirely new approaches. Elysium has a particularly interesting method, using biomimicry to create electric tendons that stretch and contract on demand, similar to the tendons that move fingers on human hands. Optimizing existing motors and entirely novel designs will reduce costs and dramatically improve the capability of these robots.


Just as open-sourced designs of four-legged robots such as Stanford’s Doggo and MIT’s Cheetah have built out a supply chain of inexpensive actuators and inexpensive dog robots in China, we anticipate that once one large company starts building out a supply chain for parts optimized for humaniform robots, the cost to build them will dramatically reduce the cost of humaniform machines, and finally make them economically viable.


The confluence of cheaper actuators, inexpensive compute, and AI enabling much easier training and many more use cases gives us conviction that humaniform robots are no longer science fiction. When building the robot is inexpensive, perhaps it then does make sense to have just one general-purpose robot that does a variety of tasks. With reasonable pricing, Asimov’s thesis on robot economics will come true, and we will live in a world surrounded by robots that look, work, and move like us.

bottom of page