The Lem Objection

Are Human-Like Robots Feasible? Maybe. Are They Worth the Trouble? Maybe Not.

The other day I got an announcement from Oregon State University that it has captured the Guinness World speed record for a robot running the 100-meter dash — just under 25 seconds. This is impressive (no falls! the press release said). It is also a lot slower than the human record (less than 9.6 seconds). Which is a reminder that many actions that are easy for humans are fiendishly difficult for machines. Running (or walking) on two feet (with no falling over!) is one of those. Which is why some knowledgeable people (for example, Gary Marcus) are skeptical of tomorrow's Tesla demo of a human-shaped robot.

"Cassie" is Oregon State University's sprinting champ

But then, why do robots need to run around on two feet? Why not give them a more stable four, as "robot dogs" have? Or five legs, or 20, or a nice set of wheels? Many successful robots take the path I wrote about last year: Give robots those parts of human work that are inherently "robot friendly" (for example, "wander around this room, vacuuming, over and over") and not those parts that aren't congenial to machines (for example, "pick up the toys and scattered clothes and pencils in this room and decide what goes where before you vacuum."

The design work here is about assembling available hardware and software into something to do the defined task. That's easier than trying to solve the problems involved in imitating a human body, much less a human mind. And being easier, it seems like a better use of limited robot-research resources. I'd rather have a robot drone that can fly into collapsed buildings looking for earthquake survivors than a robot that can play King Lear. 

Among the still-unsolved problems (things that are easy for humans but still not mastered by any robot) are human-level use of feet, legs, arms and hands (grasping and fine manipulations especially). To say nothing of a human-level facial expressions and movements. 

The best human-like robot bodies in the world won't fool anyone for long. People who engage with such devices aren't entering into a world where the difference between human and machine has been erased. They're pretending to. For fun or curiosity, they're playing along, as you might if a magician asked you to get onstage and help with her next trick. 

Add to these limitations of the robot body, there is the very wide gap between what AI can do and what a human mind can. We are a very long way from AGI (artificial general intelligence) that can hear "robot, can you tidy up before the Smiths get here?" and understand all the implicit information in it ("can you?" = an order; "tidy up" means put things in order and clean, but not to space-capsule-sterile standards; the Smiths have certain expectations that aren't the same as the Jones', and so on). 

For these reasons some people argue that a truly human-like robot is impossible, no matter how much we improve our hardware and software. I don't know if they are right, and I don't much care. 

The crucial question raised by Tesla's promise (and its spiritual cousins the geminoids, Sophia, Jia Jia, latex and silicone sex robots, among others) is not whether we can build an ersatz human, but whether we should. Since it's extremely hard to make a robot with human-level abilities, why not concentrate AI and robotics research on more attainable and useful goals?  Let's make robots that can fold fitted sheets, pick up drywall, dig foundations, lift patients into hospital beds, without giving them faces, hands or voices. That seems like a better use of limited robot-research resources. 

The answer to this, the pro-humanoid-robot case, is two fold. 

The first claim is that we want robots that will help us in the world we inhabit — a world of stairs, doors, desks, stoves, bathtubs, cars, assembly lines, stores and millions of other things that were built for hands, arms, legs and feet. To be useful, robots will need to get by in environments built for our bodies, using tools built for us. So they should be shaped like us. This part of Tesla's rationale for the robot it has promised to show off tomorrow: It wants robots that can work in factories built for human workers, so they need to be human shaped. Ditto for other uses that Tesla's CEO Elon Musk envisions: mowing the lawn (at least with a mower made for a person), making dinner, holding a conversation as if it were a friend.

It's odd hearing this argument from people who say their robots are going to change the world. If near-future robots are revolutionary devices that are going remake every aspect of life, then why are the bound to fit into homes, workplaces and streets built in the 20th century? The automobile changed the streetscape of cities and the landscape of the countryside. Why shouldn't robots? As this kitchen robot prototype from Toyota shows, it's quite possible to have a human-friendly kitchen with a non-human-shaped robot in it. Robots can climb stairs, screw in a part, and make burgers without looking anything like a person.

The second argument for human-like robots is harder to dispute. It's this: The more a robot is involved in our lives, the more information we'll have to exchange with it. That doesn't just include explicit instructions ("put everything in the garage, and even though I said 'everything' I don't mean to include the dog, the food or the lawn"). It also includes intentions, implicit information, degrees of understanding. We could spend a lot of effort to create a new way for a robot to indicate how well it understands what it's told. But it's more effective to give it a face and let it show a quizzical, half-certain expression in response to what we've just said. 

Well, true. If I want to know how confused my robot is, or how concerned it is about my health, a facial expression will tell me more, more quickly, than any text or flashing light. But at what price? 

One cost will likely be a diminishment in the value we place on the gestures that robots imitate. Once I have seen a machine perform a delicate expression of interest, apprehension and affection, I can't see that expression as the uniquely human product of a mind. After all, I've just seen a machine do it! That the machine doesn't look really human only adds to the devaluing effect. 

As Frank Pasquale puts it in The New Laws of Robotics: Defending Human Expertise in the Age of AI, "when the counterfeiting of money reaches a critical mass, genuine currency loses value. Much the same fate lies in store for human relationships in societies that allow machines to freely mimic the emotions, speech, and appearance of humans." 

    Second, there is a fundamental doubt about this sort of project, that I think of as the Lem Objection, because I found it best put in a novel by the great 20th-century SF writer Stanislaw Lem. It is this: The closer we get to reproducing a human being in robotics, the less we gain from this "plagiarism" of nature. The value we get from robots is precisely their inhumanity — their lack of emotion, their inability to experience boredom or a yearning for new horizons. Why would we want to give that up, to create an artificial version of our own flaws? Making an artificial person, as Lem says, "would be like finally building, after colossal expenditures and theoretical work, a factory for making spinach or artichokes that were capable of photosynthesis—like any plant—and which in no way differed from real spinach and artichokes except that they were inedible."