Roboticists at Carnegie Mellon University have created a training process that enables robots with abilities limited to tasks such as opening doors or drawers, for instance, to improve as it goes.

According to the researchers, the robots were trained to perform tasks under real-world conditions. An adaptive learning approach was developed that enabled a robot with a limited knowledge base to learn via hands-on experience.

Source: Carnegie Mellon University

To accomplish this, researchers constructed a four-wheeled robot featuring a single arm and grasp hand unit using off-the-shelf components. The only task that the robot was designed to perform was to approach a door or drawer and then use its clasper to grip, turn or push open the door or drawer. Initially, the researchers demonstrated to the robot how to manipulate doorknobs to gain entry to a room or building.

During an outdoor trial of the system, the robot attempted to open a door. The team observed that if the robot had already encountered a particular knob type, it could easily open the door. Conversely, if the robot encountered an unfamiliar knob type, the robot would attempt to employ its knowledge of other doorknobs to open the door.

Eventually — up to about half an hour — the researchers found that the robot typically figured out how to open the door or drawer, demonstrating an estimated 95% success rate.

An article detailing the research, “Adaptive Mobile Manipulation for Articulated Objects In the Open World,” appears in the journal arXiv.

For more on the process, watch the accompanying video that appears courtesy of Carnegie Mellon University.