Toys have a lot to learn from robotics, said Hanns Tappeiner, CEO of Anki, the company that’s using artificial intelligence (AI) to make next-generation toys that are aware and able to interact with their surroundings. Anki’s $150 race car set, Overdrive, which has cars that can see the track, may be this Christmas’ big hit. It is the second best-selling toy on Amazon.
Your vision of a post–Norman Rockwell Christmas has kids in pajamas with pistol grips speeding toy race cars on endless figure 8s, a game that gets old before January and ends up in the closet. The new race car sets will, however, blow you away. With Overdrive, the magnetic snap-together track has no slot. Cars take a lap to learn the track and will remember it. They can attack each other with “lasers” as well as other weapons drawn from a small arsenal, all of which you can control with your smartphone. Should you not be able to arrange a play date, there is a single-player mode.
Although video games have evolved from their humble beginning to a present form that includes more complex interaction and graphics that are virtually unrecognizable to Pong’s creators, the way we interact with toys has changed very little. The advances being made in robotics and machine learning, if applied to toys, is where Tappeiner sees an opportunity.
Design Smart Toys, Not Child’s Play
Tappeiner is a product of Carnegie Mellon University’s doctoral program, an institution perhaps without equal for its robotics research and development. With Overdrive successful, Anki is looking for its next hit, and hopes to do it with Cozmo.
Cozmo is a white, fist sized machine/computer/toy on wheels that Anki would like you to “adopt” -- as you would a puppy.
If you start with a dumb toy, such as toy race cars of yesteryear, any intelligence added to it would be and improvement, moving it up on the intelligence scale. But making starting a concept of a pet and trying to get those qualities into a toy is going in the opposite direction.
Programming even the simplest human interactions were challenging. Although babies learn how to make eye contact on their own, a machine can only do so with a few thousand lines of code. “We were able to call up facial recognition, a 9,000-line algorithm, with just two lines of code,” said Tappeiner.
Cozmo is shown sleeping. Its LED “eyes” are half shut, and it plays a snoring sound. “Awww,” goes the audience. But it’s during its awake state where it must try to act like a pet — not an easy job for a machine that looks more like a miniature earth mover than a puppy.
Baby Steps
Normal human or pet behavior, learned early and taken for granted, is behavior that must be intricately coded and inserted in order for toys to even have the slightest chance of endearing themselves to kids. But attributes that make all humans instantly like the puppies do not compute easily for Spock-like engineers.
“We found Cozmo had to ‘look’ at its owner every eight seconds, on average,” said Tappeiner, citing research in which babies that were not able to fixate on their mother’s face were diagnosed with autism.
Cozmo has been trained to play a game in which it smacks a sensor, trying to beat its owner. It’s fast and wins a lot. When it loses, it gets mad and has a little tantrum, smashing the sensor cube. It draws laughter.
That bit of human weakness is enough to make it relatable.
Perhaps the hardest lesson for brilliant engineers to learn is that for us to like our machines, it’s not that they have to be stronger, smarter, faster … They have to be purposely flawed. An engineer’s mindset, education and training combine to make robots that can spot weld without breaks on an assembly line or can beat the best of us at Jeopardy! All of that has only succeeded in making us like robots less, with a wary appreciation at best and a dread of being annihilated by our own creations at worst.