Humanoid robots have caught the public imagination for over 100 years. Ever since the word ‘robot’ was coined into the English language, in fact, via Czech playwright Karel Čapek’s 1920 drama, R.U.R (Rossumovi Univerzální Roboti, or Rossum’s Universal Robots). The word came from the Czech or Slavic ‘robota’ meaning forced labour.
But despite the artificial humans conjured by that play, and the generations of fictional tin men that followed, humanoid robots have rarely done anything useful in the real world. Robots have built cars, of course, and performed countless other automated tasks in factories, electronics labs, and extreme environments. But few, if any, of them have been humanoid.
So, humanoid robots’ success has been muted at best. Toys have proliferated, of course, while robots like Honda’s ASIMO and Tesla’s Optimus have been brand leaders for their corporate masters – the first convincingly twenty years ago, the second less so today.
Elsewhere, humanoids like Aldebaran’s NAO and Pepper have been entertaining but limited public relations tools, due to poor intelligence and few useful applications. Meanwhile, Boston Dynamics’ hydraulic machines, like Atlas, have at least performed one critical function: expanding our awareness of what robots can do – running, jumping, somersaulting, and climbing. Impressive. But beyond that, what are they for?
One of the challenges is that a century of science fiction has raised expectations of what interacting with humanoids should be like: super-intelligent mecha companions. But the reality has always been disappointing: simple responses to trigger phrases, or pre-programmed routines that swiftly become boring – however impressive a robot’s design or engineering may be.
But all that is changing fast with the rise of chatbots like OpenAI’s epochal ChatGPT. Put a large language model (LLM) into a humanoid robot, as the UK’s Engineered Arts has done with its Ameca android, and you can at least simulate intelligence. A parlour trick, but an impressive one, when combined with slick design and engineering. But again, what useful jobs can it do? Why not just employ a person?
So, how close are we to a viable, general-purpose humanoid? One that can not only do something useful (beyond being an entertaining distraction, like Ameca), but also demonstrate a genuine ROI? In short, where is the humanoid robot that can labour alongside flesh-and-blood workers safely and predictably in real-world environments?
Very close, according to Austin, Texas-based engineering consultancy Apptronik Inc, a spin-out from the Human Centered Robotics Lab at the University of Texas.
In fact, the 85-strong company is launching its Apollo electric humanoid today [23 August]. The multifunctional, human-sized ‘cobot’ is designed to perform repetitive manual tasks, such as shifting boxes and crates in factories, warehouses, and large retail environments.
Apollo’s target function is gross manipulation – repeatable, whole-body motor tasks like bending down, picking up items, and moving them – rather than the complex manual tasks that Sanctuary AI is aiming at with its own general-purpose machine, Phoenix. (The question ‘Why does a robot need to be humanoid?’ is answered by the fact it will operate in environments designed by, and for, human workers.)
Apptronik’s claim that Apollo is a viable product today, at a cost equivalent to a high-end family car, is a bold one in a robot world dominated by eye-catching designs, solid engineering, but unconvincing applications. That the company has reached this point so quickly after spinning out in 2016 is impressive.
Indeed, co-founder and CEO Jeff Cardenas tells me that Apptronik has had a hand in other robotic successes, including NASA’s multimillion-dollar Valkyrie space robot, which is being developed for work on Mars.
Two of my co-founders are [Advisor] Dr Luis Sentis, who runs the Human Centered Robotics lab, and [CTO] Dr Nick Paine, who was his first student. They were working on NASA’s Valkyrie, so you can think of Apptronik as the commercialization of the work from Luis’ lab.
Luis is known for inventing some new control theory, Whole Body Control, while he was at Stanford. He had worked with Honda's ASIMO, then came to the University of Texas back in 2010-11, where he established the new lab. Together, they built a variety of different robots, and a lot of that work culminated in Valkyrie, one of the most advanced electric humanoids that had been built at that time.
Boston Dynamics had Atlas, its hydraulic robot, but Valkyrie was electric [and therefore safer around humans]. Now, the thesis of this company is around some key ideas. One is that robots have to become more versatile, which was the premise of the DARPA Robotics Challenge: increasing versatility and building more general-purpose machines.
But Valkyrie was 300 pounds, six foot two, and millions of dollars. And what we were chasing was building a commercial humanoid, at a target price less than many cars. So, how do you produce a robust, affordable, scalable, capable, general-purpose robot? That is the focus of the company.”
The result is 5’ 8” Apollo, a modular, and therefore upgradeable, humanoid co-worker. Swappable battery life is currently four hours, the robot has a simple, intuitive onboard interface, plus a mission-building remote operating system that allows human managers to set tasks for robots to carry out autonomously.
Apollo is the culmination of the last seven-plus years of work, both in the lab and as a company. This is our first true commercial product, so we're excited to get it out of the lab and into the real world.
The company’s mantra is ‘Not Man vs machine, but Man plus machine’ (in 2023, ‘Human plus machine’ would be better), while competitors – and partners – like Sancturary AI maintain their rival products are designed to do the jobs that humans don’t want to do at all.
So, who are Apptronik’s likely customers? Which industries would benefit from having “friendly and collaborative” robot co-workers? Cardenas says:
Logistics, retail, and manufacturing, all around the same basic applications.
Long term, the idea of a general-purpose robot is a single machine that can do a whole range of different tasks. But what we've learned by deploying these systems, and building them with customers, is that go-to market for humanoids is gated by technical feasibility.
Meaning that, yes, theoretically, these robots can do anything at some point. But what can they, technically, achieve today? And in a big enough market where we can build a viable business around it and get started now?
So, the path towards general purpose looks more like multipurpose initially, if we can get a robot that can perform three different tasks during a day. But that is already a game-changer for our customers: one task in the morning, another by lunch, and a third at the end of the day. That changes everything that they know about robotic automation.
There are thousands of robots that do one or two things. But the goal with Apollo is to build one robot that can do thousands. However, the path to real general purpose is, first, multipurpose. And before multipurpose, the robot has to do at least one task reliably, at scale and at rate, with a positive ROI.
And that is box handling – gross manipulation. The customers we're engaged with all have similar profiles, but we're going to start in manufacturing and distribution, in crate- and box-handling. Initial applications such as picking to and from a conveyor, or to and from a pallet. Then we're looking at things like truck loading and unloading.”
At present, Apollo can lift 55 pounds. But for such simple tasks, why build a robot that has legs – risking instability – rather than wheels? He says:
Apollo is modular at the torso, so you can mount it to wheels. But we believe that legs will win the day for robots in the long term.
The advantage of legs is that you take up, literally, a very small footprint, but you can bend at the knee to reach the ground, or you can reach up high. If you want to lift something heavy, you just move the feet further apart.
But if you want those same advantages with a wheel base, you end up with a very big, wheeled platform that typically has to be heavy. Plus, the ability to seamlessly integrate with existing workflows designed for humans means that legs, ultimately, will win the day at scale.
For us, the exciting thing is we have a technology that works today. And our goal is for it to operate with full, supervised autonomy.”
But the simplest way to control a robot – as we know from a century of sci-fi – is to tell it, verbally, what to do. With large-language models entering the mainstream, how close are we to building an intelligent, general-purpose machine that can understand complex instructions?
Talking to the robot is going to be an important part of the interface in the future. But we don't want to rely on that yet, because it’s a big challenge and it means more complicated robots.
So, we've tried to build on the software side initially, to create an easy interface to abstract a lot of the complexity away. To make it easy to tell the robot what to do, and for it to verify the task.
But yes, eventually you’ll be able to say ‘pick up those items and move them there’ just like you tell a human worker. And the robot will need to validate its performance: ‘Am I doing what you want?’.
All that will be important, and it’s become more feasible in the last 12 months with LLMs. With our system that could be a software upgrade away.”
A promising development in robots’ slow march out of the pages of sci-fi and into the real world. It has taken 100 years to reach this point, but the collision of several technologies may finally make those visions a reality.
But which visions exactly? If science fiction teaches anything, it is ‘choose your future wisely’. And if business teaches anything, it is that a good product alone is never enough; it also needs to do something useful, affordably. As another robotics CEO once told me, “Nobody makes money from selling robots. You make money by selling useful services.”
So, will Apollo be useful enough? We are about to find out.