A team of American roboticists at Carnegie Mellon University in Pittsburgh has developed a compact wheeled robot capable of opening doors and drawers without step-by-step instructions. The device learns by trial and error, updating its actions in real time as it encounters new hardware. The project details appeared in a study shared through open-access portals and arXiv-style publications.
Engineers outfitted the unit with a versatile manipulator and programmed it with a library of handle shapes and lock mechanisms. After initialization, the robot was placed in a testing zone where it faced a variety of doors and drawers to try to operate.
Experiments demonstrated that when the robot could identify the type of latch or handle, it quickly employed the most relevant approach based on its accumulated knowledge. When faced with an unfamiliar configuration, the system drew on broad principles learned from similar objects to deduce a viable action.
Researchers observed that the system could successfully open about 95% of doors and containers after sufficient trials. In some scenarios, it required up to half an hour for the robot to determine the correct sequence of movements to reach a goal.
Earlier efforts in robotics included a model designed to navigate and clean an unfamiliar living space, illustrating the broader aim of letting machines adapt to new environments with minimal human guidance.