The internet and VCs love humanoid robots. Will they return the love?
Cleaning a floor might be accomplished using a complex humanoid robot with say 75 degrees of freedom to push a manual vacuum. Or the job can be done with a profoundly simpler robot designed for just that purpose.
The current fascination with “general purpose” humanoid robots—equally proficient at rocking the baby to sleep or jackhammering concrete to rubble—is not new. The notion was popular when I first became captivated by robots in the early 1980s. Advocates declaimed: We live in a world made for humans. Every tool and every process is optimized for people. Thus, humanoid robots are ideal drop-in replacements for people. And it obviously makes more sense to build one robot able to perform any task rather than engineering dozens of purpose-built robots that each do just one thing.
Robot Versus Robot
Transforming the idea of humanoid robots into functioning hardware proved excruciatingly difficult. Honda was among the first to take up the challenge. And, after 14 years of careful work they debuted humanoid Asimo in 2000. The clever robot could walk up and down steps, push a cart, and even kick a ball. Asimo never reached the marketplace but if it had, Honda estimated it would have cost $2.5M. By contrast, Roomba, a robot that “only” cleaned floors, launched in 2002 with a price tag of $200.
Twenty-five years ago, the difficulty and cost of building a humanoid robot to perform a task dwarfed the challenge of creating a purpose-built robot to accomplish the same thing. (And building single-purpose robots was plenty hard! See DancingwithRoomba.com.) Had the Roomba team chosen to solve consumer floor cleaning by creating a humanoid robot to wield a manual vacuum cleaner, our project might still be underway. As it turned out, Asimo never cleaned a consumer’s floor, while Roomba has cleaned tens of millions. We made the right choice when we built a robot that did one thing rather than a robot that aspired to do everything. But technology marches on. Should inventors and investors make the same choice today?
General-Purpose versus Purpose-Built
The tradeoff between general-purpose and purpose-built solutions in a particular domain changes with time. Consider this example. Suppose that in the year 1951 you’d been tasked with making a light blink automatically. How might you accomplish this? Two options present themselves. You could, 1) Use a general-purpose computer and specialize it to the light-blinking task by writing software, or 2) Create a purpose-built circuit whose hardwired, single-function was to blink the light.
The general-purpose solution is attractive because it lets you easily change the blink rate or blink pattern of your light simply by typing. To achieve this feat in 1951 you could purchase the newly released UNIVAC, the first general-purpose computer, connect an output line to your lightbulb, and then write a simple program. The computing machinery would weigh eight tons, draw 125 kW of power, and cost around $1M.
The alternative, a purpose-build circuit, would probably include a relay, a capacitor or two, and one or more resistors. It might take half a day to design and solder the circuit together and once completed, you could easily hold it in one hand. The total cost would have been a few dollars. But, if you wanted to change the blinking parameters, you’d need to power up your soldering iron, remove some components, and swap in others with different values.
In 1951 the purpose-built circuit would have won hands down. Mid-way through the twentieth century, the versatility and convenience of a general-purpose solution to an electronic control task couldn’t justify the then astronomical cost.
Fast forward to 2002 where we find the Roomba team confronted with their own light-blinking problem. We needed to design a “virtual wall” transmitter to keep Roomba from wandering into places where it was not wanted. Our virtual wall unit projected a narrow light beam. When Roomba encountered the beam, it turned away. Roomba distinguished the light beam emitted by the virtual wall transmitter from all the other distracting lights in the room by its blinking pattern and rate.
The idea of designing a single-purpose circuit to blink our light (an infrared LED) was at most a passing thought. We settled on using a general-purpose computer almost immediately. The computer we chose was an especially simple one—a tiny microprocessor that cost about 11 cents.
In the decades between the launches of UNIVAC and Roomba, a once prohibitively expensive general-purpose solution for an electronic problem fell in price by a factor of 10 million or so. Based on a mass-produced component, the versatility, ease of implementation, and reliability of the general-purpose solution made it the best choice by far. Even if heroic efforts had enabled us to design a bespoke solution for slightly less than 11 cents, we’d still have chosen the microprocessor for those reasons.
Ready for Prime Time?
Contemporary advocates for humanoid robots argue that their time has arrived. The amazing advancements in AI now convulsing the internet and society, will transform robots as well, they say, making humanoids practical and commonplace. VCs have bet tens of billions of investment dollars that humanoid robots have achieved 11-cent-microprocessor status—that it now (or very soon will) make more sense to specialize a complex, general system to perform a particular task through training or programming than to design a bespoke system to accomplish that task.
Economical general-purposeness is a very high bar. Currently marketed humanoid robots are priced roughly between $20,000 and $100,000. So, compared to Asimo, the price for a humanoid has fallen by a factor between 25 and 125. That’s a far cry from the factor of 10 million reduction Moore’s Law granted computational electronics. But sadly, motors, batteries, and steel don’t scale the same way as transistors.
Still, at the most optimistic price, a truly general-purpose humanoid robot able to perform any task a human can do would make economic sense.[1] So, are we there yet?
No. The current crop of humanoid robots falls short in several ways. One is dexterity. People are great at manipulation tasks. We can, for example, examine tactilely unseen coins in our pockets and pick out the desired one. A variety of types of sensors built into our fingers enables such feats. Robots don’t possess the same range of sensors and research that might grant robots human-like dexterous abilities lags behind other abilities. (For a well-reasoned, evidence-based analysis of robots’ dexterity shortcomings see: https://rodneybrooks.com/why-todays-humanoids-wont-learn-dexterity/.)
Safety is another humanoid sore point. The same clever control scheme (called Zero Moment Point) that enables humanoids to walk with a natural gait can cause bad behavior under some circumstances. Should the robot slip or begin to fall, the control system may command the legs to accelerate rapidly. Flailing limbs could then injure a person standing too near the robot.
A robot that walks on two legs must actively balance. Should the computer controlling the robot crash, should a sensor fail, or should a motor break down, the robot will most likely topple over. That puts nearby people, pets, and property in danger.
Modern AI techniques have boosted the performance of intelligent systems tremendously, but they harbor a robust flaw: sometimes they make stuff up—they are commonly said to hallucinate. If the output of an AI is text or an image, the occasional gaffe may be forgiven. But the possibility of spurious physical actions makes AI-powered robots unreliable and, until proven otherwise, potentially dangerous.
The fact that AI-powered robots sometimes behave erratically does not render them unusable. But it does mean that we should use robots (humanoid and otherwise) gingerly. Robots are best employed in situations where, when they work properly, they add value, but when they fail, they do no harm. Thus, humanoid robots ought not to be trusted in every scenario where we’d otherwise employ a person.
The fundamental unreliability of even the best designed robots is why, throughout my career, I’ve strongly favored working on small robots. No matter which of their components fails or how addled their brains become, the little guys just aren’t able to do much damage.
Form Frustrates Function
The versatility of robots always fascinated me. I can’t shrink to the size of a blood cell, steer through the circulatory system, and then attack an errant blood clot. But I might build a robot that can. I can’t float with the current near the bottom of Antarctic seas collecting data for months, but recently a robot did exactly that. No army of human farmhands provisioned with delicate brushes stands ready to pollinate millions of acres of crops. But if declining numbers of insect pollinators force the issue, we might build swarms of butterfly-like robots to take their place.
We humans congratulate ourselves on being the world’s greatest generalists. Indeed we are—yet there remains a vast array of niches and critical tasks beyond our grasp. A million animal species can go where we can’t and do things we can’t easily do. For instance, should the need arise, how would people take over the dung beetle’s job? We humans are stuck with a certain scale, sensor suite, and actuator set. Robots bear no such encumbrance. Robots can be of any size and have any properties that we can imagine and then execute.
When we build robots in our image we impose our limitations on them.
Task not Tech
Electro-mechanical mechanisms, humanoid robots included, have practical constraints. These include physical size, weight, battery capacity and run time, maximum power output, payload capacity, maximum joint torque, maximum joint speed, positioning accuracy and stiffness, sensor type, and resolution along with a host of other parameters. The robot designer must make explicit choices for each of these items. The optimal choice is dictated by the task to be performed and is different for different tasks. Specialization to particular tasks occurs even among we general-purpose humans. For example, a successful linebacker is unlikely to have an equally successful second career as a jockey.
When we insist that a single mechanism perform different functions we’re forced to compromise. We may need to deliberately degrade the performance of one task to enable an antagonistic task to be done better. It does not follow that a mechanism able to perform several different tasks will do all or any of them well. For instance, attempts to build flying cars have resulted in machines that are at once poor cars and poor airplanes. And critically, a mechanism whose primary virtue is versatility is not guaranteed to prevail in the marketplace against efficient, bespoke mechanisms that make no compromises.
Sometimes multiple tasks must be done at once. A humanoid can’t wash the dishes while simultaneously folding clothes, say. And, if one robot is responsible for all tasks, then no tasks are completed when the robot breaks. These examples illustrate that satisfying customers’ actual needs requires thinking carefully and explicitly about those needs. Assuming that AI magic will tackle all problems is likely to lead to disappointment in the marketplace.
Roomba won its race because the design team valued task above technology. While other robotic groups pursued general-purposeness, exotic technology, or coolness we focused on one task. Cleaning floors at a price comparable to existing vacuums drove our every decision. It was necessary to exploit every possible feature of the domain and exclude every superfluous component in order to achieve our purpose.
This old roboticist wonders if contemporary designers are forgetting a fundamental lesson. That we are building products, not inventing the future. That products live or die by their fitness and economy for the task or tasks to which they are applied. Consider the task, then figure out what to build, not the other way around.
The danger is that in an uncritical, exuberant rush to build robots that look like us, opportunities will be lost. A generation of bespoke robots that competently and economically perform tasks we want done could be starved of capital—gobbled up instead by charismatic but not-ready-for-prime-time humanoid robots.
Let’s not do that.
Humanoid Robot Blues
Everyone loves the robots
that are shaped like me and you.
But I don’t trust those shifty things
and they’ve got me feeling blue.
Humanoid, humanoid robot blues.
Humanoid robot blues.
In contrarian isolation I just cry!
(With apologies to Heartbreak Hotel)
[1] If a humanoid robot we purchase works say, 20 hours per day, 365 days per year it would deliver 7300 hours of work annually. If we insist on a return-on-investment of two years, we’d need to recover the capital cost of the robot ($20,000) during that time. Thus, the effective hourly cost of the robot would be 20,000 / (2 * 7300) = $1.37 per hour (ignoring the cost of electricity, consumables, and so on.). That number is well below minimum wage. By this reasoning, a truly general-purpose humanoid robot could outcompete humans for any job.
