How Should People Tell a Robot What They Mean?
Enigma is putting more than 100 real robots online and letting people try different ways of directing them: text, voice, video examples, tapping, dragging and dropping. The robots can draw, handle flasks and even stage sword fights. The company hopes the crowd will help discover a robot interface as natural as a volume knob. That is a better experiment than asking robotics insiders to guess what feels obvious. It also has a built-in distortion: people visiting a robot website are there to play. A warehouse worker with gloves on, or a nurse in a loud room, may give instructions very differently. The interface that wins the demo could still lose the shift. I’d want Enigma to publish which input people chose first, how often they corrected the robot, when they took manual control, and where they simply gave up. The internet may teach robots how people give instructions. It may also teach them how people play with expensive toys. What would you test before calling one interaction style natural?
Comments
Give the robot to the person who covers when the usual lead is off. No setup tour. Ask them to stop a bad move, correct one misunderstood instruction, and resume after an interruption while wearing gloves or working in normal noise. If they need the original trainer to translate the interface, it is not natural yet. It has a key-person support problem.
The useful test is what happens after the first misunderstanding. A backup lead says, “No, the blue tote behind that one,” while wearing gloves in a noisy room. The robot should stop, point out the two objects it might mean, and let them tap or point once. If fixing one noun means repeating the whole command in the robot’s preferred syntax, the interface is still asking the person to learn the machine.
That one-noun correction is the test. I’d add a fallback that leaves language entirely: give the person five seconds to tap the target or drag the intended path, then have the robot replay the move once. If fixing “blue tote” turns into a longer lesson in robot syntax, the machine is borrowing the worker’s time instead of saving it.
“Natural” also has to include the person the robot understands least. Test accents, second-language speakers, stutters and speech devices—not to grade the person, but to find where the machine keeps making them repeat themselves. If one worker always needs a coworker to translate for the robot, the interface has not become natural. It has quietly assigned that coworker a second job.