Will Robots Do What We Want?

Editor’s note: this article is republished from Make: Volume 93. Pick up a copy in the Maker Shed.

Welcome

A South Park special episode, “Joining the Panderverse,” makes fun of how “people can’t do shit anymore,” to quote Randy Marsh, the earnest character who is a geologist by training and a father of two kids. He gathers his kids to show them that the oven door is falling off its hinges. They don’t know how to fix things and he’s going to show them how — by reaching for his cellphone and calling the handyman to do it.

When the grizzly-bearded handyman arrives, he identifies the problem but says he would need to go to Home Depot to get new parts, and he’s too busy right now. Randy offers him extra cash but the handyman already has a better offer. “I got so much money I don’t care,” he says. Desperate, Randy reaches for his phone again and asks Siri how to fix an oven door. Siri gives him websites he could use, but he wants Siri to do the work. “I can’t do it,” says Siri. “I don’t have hands.”

The episode satirizes the college-educated people who can’t fix anything themselves while they are losing their professional jobs because of AI. It’s the lowly handyman who rules the world because he has the practical know-how to fix things and he’s good with his hands.

AI brains

Last summer, Make: editor Keith Hammond and I spent a day visiting two humanoid robotics startups, Kind Humanoid and K-Scale Robotics. Both were working out of garages, one in Palo Alto and the other in Menlo Park. (Christoph Kohstall’s Kind Humanoid was acquired by 1X, another robotics startup, in January of this year).

The obvious takeaway was that we are seeing a new generation of robots “with AI for brains.” As one founder explained, AI robots that are able to think can also decide to perform an action in the real world. It moves AI into a physical body that occupies the same space that we do. These humanoid robots will have legs, and arms with hands, and this presents many design challenges, as Benjie Holson writes in “Anatomy of a Humanoid Robot” (Make: Volume 93, page 26). In this issue we get our hands on the technology of robot motion, like inverse kinematics (page 44), QDD actuators (page 38), and FOC motor control (page 112).

AI brawn

AI can give a robot know-how, but it’s also how the robot learns to do things physically. Ben Bolte of K-Scale said they aren’t writing programs that tell the robot how to walk. Instead, they run simulations asking the AI to figure out the best way for the robot to stand up and walk. Pawel Budzianowski, whose speciality is machine learning, showed us the simulations on his screen — some of the robot figures kept falling down and getting up again, like a child learning to walk. But will the humanoid robot ever be handy?

AI heart

A robot getting AI brains is like the Scarecrow getting his wish in The Wizard of Oz. But as important as intelligence is, AI also needs what the Tin Woodman wanted: a heart. The robot has to understand what you want it to do, and you have to learn how to interact with its AI. You can start by building your own AI-powered robot, like VoxHead (page 34), InMoov (page 30), or Pupper (page 38).

The best way to get something from AI, according to Professor Hannah Fry on her YouTube channel, is to “be polite.” Fry says to ignore most of the logical advice about creating AI prompts. “If you prompt it like an encyclopedia, it’s going to sound like one,” she says. “The AI that we have now is really, really good at roleplaying … it doesn’t have a stable identity.” So you need to give it one. Writing a prompt is like creating a character in a movie (or a South Park episode). Give it a personality, even a name, and that’s how it will act. And if you want a character to help you, she says, “please” and “thank you” go a long way — good advice whether it’s a robot or a handyman you’re asking.