Berkeley's two-armed robot hints at a new future for warehouses

Pick up a glass of water, then lift a fork: Without thinking, you chose the best way to grasp each object. Researchers at UC Berkeley have developed a robot that makes the same calculation, choosing on the fly whether to grab an object with pincers or lift it with a suction cup.

Why it matters: Reliable robot grabbers are the just-out-of-reach holy grail for e-commerce outfits like Amazon and Walmart, who still rely mainly on human hands for the job. Smart picker-uppers would clear a serious bottleneck in shipping and could change the nature of warehouses entirely.

How it works: Berkeley's two-armed robot, seen in the video clip above, first considers the contents of a bin and calculates each arm's probability of picking up an object.

• Its suction cup is good at grabbing smooth, flat objects like boxes, but bad at porous surfaces like on a stuffed animal. The pincers, on the other hand, are best with small, odd-shaped items.
• The system learned its pick-up prowess not from actual practice, but from millions of simulated grasps on more than 1,600 3D objects. In every simulation, small details were randomized, which taught the robot to deal with real-world uncertainty.
• The bot can pick up objects 95% of the time, at about 300 successful pickups per hour, its creators write in a paper published today in Science Robotics.

Warehouse robots that can move around merchandise are highly sought after. Amazon is reportedly working on its own "picker" robots, as are several robotics companies.

• Whoever succeeds could kick off a total reimagining of warehouse distribution, says Ken Goldberg, a Berkeley robotics professor and co-author of the paper.
• "You could have very dense warehouses where you could have these bins and robots in really tight quarters," says Goldberg. Think small, vertical warehouses throughout Manhattan, rather than a few huge warehouses mostly outside the city.

We've reported on a JD.com warehouse in China that only employs four humans — but it only handles boxes, not loose items. The difficulty of grabbing never-before-seen items — trivial for a human — means totally automated warehouses are still a ways off.

• One limitation of the Berkeley bot is that it can't change its plans once it begins moving to pick up an object, and therefore can't react to its environment — like items resettling in a jostled bin — says Jonas Schneider, head robotics engineer at OpenAI, who was not involved in this research.
• Jeff Mahler, a Berkeley researcher and the paper's lead author, says the system only "sees" the bin once every 12 seconds, causing more problems if things shift around.

What's next: Mahler says robots will eventually need to place objects in precise orientations — on a conveyor belt used in a factory, say — and to tightly pack a shipping box. Perhaps most valuable would be a bot that can rummage through a box to find a specific item.