Robot Hands Can Now Grab Things Without Knowing Exactly Where They Are

Robot hands have a reputation problem. In labs they can do astonishing things, but put one in a messy kitchen or a cluttered warehouse and it often fumbles. The reason is that most robot grasping software needs a precise 3D model of the object and exactly where it sits. In the real world, you rarely have that. Objects get moved, lighting changes, and no two apples are identical.

A team of researchers took a different approach. Rather than demanding accuracy, they designed a closing motion that is naturally forgiving. Their planner uses four simple ideas: a rough outline of the object instead of a detailed model, grip styles borrowed from how humans actually hold things, a focus on the object itself rather than a fixed set of contact points, and fingers that are allowed to slide and settle into place. Think of how you reach for a door handle without looking — your hand adjusts on contact.

The results are encouraging. Using a Shadow Dexterous Hand (a well-known robot hand with five fingers), the planned motion worked even when the object's size was off by about a centimeter and its position was off by several centimeters and tens of degrees. It beat a leading data-driven planner in 25 of 27 tested scenarios. In a full autonomous test, it successfully grasped 82.5% of 80 everyday objects. And crucially, it did all this without any feedback — no cameras correcting it mid-motion.

The bigger idea is that robustness can be built into the movement itself, not just bolted on afterward as sensors and corrections. That's a meaningful shift for anyone hoping robots will eventually handle dishes, packages, or groceries. The team's next step is combining this forgiving motion with live feedback on a physical hand. If that works, robots that cope with mess and uncertainty get one step closer to everyday life.