A robotics team has taught a robotic arm hundreds of distinct ways to open the same drawer or cabinet door.
Researchers built QDTraj, a system that uses Quality-Diversity algorithms and sparse-reward exploration to generate low-level trajectory primitives for manipulating articulated objects like hinges and sliders. Instead of training a robot on one "correct" motion, the method searches for many different high-performing paths to the same goal. In tests across 30 object articulations from the PartNetMobility dataset, QDTraj produced an average of 704 distinct trajectories per task, at least five times more diverse than the methods it was compared against. The team validated the approach in simulation, then deployed the resulting trajectories on a real robot.
That matters because a single scripted motion breaks the moment a real kitchen does not match the training setup, whether that means a stuck hinge, a blocked drawer, or an unexpected angle. A robot with hundreds of viable ways to do the same task can pick whichever one still works when its first choice does not. It is a small but concrete step toward robots that handle the mess of real homes rather than a tidy lab bench.
Still, opening cabinets and drawers is a narrow slice of household chores, and PartNetMobility's articulations are simulated stand-ins for the real thing. The gap between five-times-more-diverse and actually useful in someone's kitchen remains untested.