Here’s something disturbing: a video released by Cornell University showing a robot wielding a knife (the money shot, as it were, is above—scroll down for the full clip). The knife’s point comes perilously close to the torso of a researcher, who responds as most humans might to a robot carelessly swinging around a naked blade, and lurches away. Your attention, it’s safe to say, has been captured.
Here’s something more disturbing: Roboticists are now using stagecraft, with a touch of live-action roleplaying, to capture your attention. The knife is blunted, and that scripted, last-second dodge is strictly for your benefit. But hey, it sure looks like a robot almost gutted someone, right?
The experiment in question is actually an interesting one, demonstrating the kind of training that could turn fumbling robots into more effective everyday servants, both in the workplace and at home. The bot in the video, Rethink Robotics’ Baxter, is already designed to perform repetitive tasks by being taught—the user can physically move the robot’s arms and manipulators, the equivalent of a professional golf instructor or overly flirtatious pool shark guiding another human’s limbs through the necessary motions. The robot records and repeats the sequence, and, in Baxter’s case, makes minor adjustments and adaptations, recognizing when a manipulator should be rotated to better pick up an off-angle or tipped-over component.
But if Baxter is ringing up a customer—a purely hypothetical scenario, since that’s not one of the bot’s out-of-the-box uses—and grasps a knife, it doesn’t have a sense of context, propriety, or manslaughter. It’ll blithely swing that blade close enough to terrify a human, and possibly set off the first of many robot-lynching riots to come. And while Baxter’s masters could try to maneuver it through separate grasp and carry sequences for every type of object or product in the store, the team at Cornell has a better idea.
They call it a “coactive learning technique,” which essentially means teaching through nudging. Instead of re-recording a specific sequence for handling knives around artery-filled humans, the user adjusts Baxter’s movements as they take place—turning the knife, for example, away from the customer. Subsequent adjustments have a cumulative effect, fine-tuning the proper way to ring up cutlery without exacerbating the already strained relations between people and their machines. More interesting still, the manhandling applied to one Baxter cashier, and the adaptive response by its coactive algorithms, could be shared to its fellow robot employees. The knife-happy hive mind will have learned its lesson.
But as assistant professor of computer science Ashutosh Saxena admits, the knife video is a bit of a gag. Naked blades are pretty rare in retail, and aren’t an urgent problem of human-robotic interaction seeking a solution. There is value, however, in learning to handle objects within the context of a given environment. “Even taking out a pen—it is so obvious to us not to put a pen close to the eye of a human, but that is not obvious to a robot,” say Saxena, whose graduate student, Ashesh Jain, led the experiments. “If you had a robot serving you coffee, if the coffee is hot, you don’t want it to be over your laptop, or over your lap. That could be dangerous.”
Along with safer handling of dull-edged knives, the Cornell researchers taught Baxter to carry cartons full of eggs low enough to avoid or limit breakage, should they slip out of its manipulator, while cereal boxes could be tossed around with relative abandon. And in a companion set of experiments, a different two-armed robot—Willow Garage’s PR2—grasped and carried objects through a household environment. The full results will be released at a robotics conference this December, but the PR2 was trained, for example, not to slosh a glass full of liquid over a laptop, and to carry a flower vase more carefully than sturdier items.
The notion of robots replacing cashiers is, to be clear, either silly, misguided, or all of the above—as Farhad Manjoo pointed out in the Wall Street Journal, even automated checkout systems are notoriously poor performers. Adding a pair of arms, no matter how algorithmically-gifted, might not turn an inherently human pursuit into an efficient, mechanized one. But for domestic robots to become a reality, cautiously fetching food-topped plates for their elderly human overlords, and ensuring that knives aren’t about to tip off of dirty dishes during retrieval, the ability to handle objects within the proper physical and social context could be a key selling point.