Deep Dives with Iman

Robot hands have become remarkably good. Iman Mossavat talks with Dr. Kushal Kedia, a postdoc at MIT, who thinks we are wasting most of what they can do.

Kushal’s diagnosis is that hardware has jumped ahead of software. A robot taught by demonstration can only be as good as the person puppeteering it, and a human is a master of their own body, not of a five-fingered hand with 22 joints. He wants “the max juice” out of the hardware, meaning letting the robot find things nobody showed it.

Play2Perfect, a paper Kushal co-authored with colleagues, starts with play. Before the hand is trained to insert a peg or screw something in, it plays with random objects, a bit like a baby, until it controls its fingers. The precise assembly then takes two to five hours to learn, with no hand-engineered reward.

Now the uncomfortable part. A robot that beats its teacher will sometimes make a Move 37, a move that looks like a blunder until you understand it. Kushal wants that, but legible to the people standing next to the robot.

A robot video rarely shows whether a skill works every time or only on the one attempt in a hundred that happened to work. Jens Kober says this is one reason demos say little about the reliability factories need.

Jens is a professor at the Learning and Interactive Robots Lab at the University of Stuttgart, after almost a decade at TU Delft. For about ten years his research has focused on letting people who need not be robotics experts coach a robot: they demonstrate a task, watch it try, then correct or rate parts of the movement.

In conversation with Iman Mossavat, he makes three points. Simulation is no guarantee of success on hardware, because contacts, friction and deformable objects are hard to model. Reward design can backfire: a robot rewarded for minimising the distance between ball and cup in a ball-in-cup game ends up hitting the cup from below, and a gardening robot rewarded for removing weeds could remove every plant. Today’s generalist models reach roughly 60 to 80 percent success, against a typical 95 percent or more for commercial use, he says.

His work keeps a human coach in the loop and adds robots that estimate their own uncertainty and ask for help.