This post is also available in: עברית (Hebrew)
The world’s leading industrial robot maker, Fanuc, is working on an incredibly clever robot. The Japanese company has created a robot that learns how to perform a task overnight, with little to no human intervention.
Industrial robots are replacing humans in tedious jobs, and for good reason. They are quick, precise, and they never get tired. The one major drawback is that for every task they perform, someone had to programme the robot to do it. This process is time consuming, difficult, and limiting. For this reason, robots have until now been used in tightly controlled environments with rigid manufacturing procedures.
Fanuc’s robot is about to change everything. The company is using a technique called deep reinforcement learning so its robot can learn to do a task over time. The robot attempts a task repeatedly while filming itself doing so. Each success or failure is recorded and analysed, and the next attempt takes these previous lessons into account to refine a deep learning model.
“After eight hours or so it gets to 90 percent accuracy or above, which is almost the same as if an expert were to program it,” says Shohei Hido, chief researcher at Tokyo-based Preferred Networks, who programmed the self-teaching robot. “It works overnight; the next morning it is tuned.”
According to Hido, robot learning could be sped up even more if several machines learnt in tandem, sharing their experiences along the way. Nine robots working in tandem for one hour could accomplish what would take one robot nine hours, greatly speeding up the process. “Our project is oriented to distributed learning,” Hido says. “You can imagine hundreds of factory robots sharing information.”