When it encounters a new object it has to pick up, Tellex’s robot first scans it from different directions with infrared sensors. When it has a sense of the object’s shape, it mathematically tries out a number of different ways its hands could grip it, optimizing as it goes. Once happy, Baxter tries out the optimal grip by attempting to pick up the object, and usually this goes quite well: Picking things up is apparently 75% more reliable this way than using the machine’s basic routines (pdf).

The other clever bit is the networking, which means many Baxters working in concert could teach each all the others, quickly building a vast catalog of grip knowledge.

Gripping things without dropping them is a critical task for robots in many lines of work, from Amazon’s warehouses to electronics factories. Recent breakthroughs in robotics have even seen the arrival of Tally, a robot can that stalk the aisles of a supermarket and reliably audit stock. Thanks to Tellex’s work, future Tallys may even be able to restock shelves too, gripping all manner of trickily-shaped cans, boxes and bags of stuff.

📬 Sign up for the Daily Brief

Our free, fast, and fun briefing on the global economy, delivered every weekday morning.