Researchers from the University of Michigan are working on making robots more functional and capable in the real world. Researchers working on the study say the models robots use to perform tasks typically work well in a structured environment inside the laboratory. However, when the robots are used outside the laboratory, some of the most sophisticated models prove inadequate in certain situations. Situations that prove particularly difficult for robots to perform are working with soft materials like rope or cloth.
Researchers at the University of Michigan have created a way for robots to predict when they can’t trust their operating models and to recover when they find the model governing their control lacking. Essentially, the team is looking to try and teach the robot to make do with what it has, according to robotics Ph.D. student Peter Mitrano. The goal of the researchers is to have the robot be able to pick things up and move them around without knowing the physics or geometry of everything.
They created a simple model of a rope’s dynamics while it moves around an open space. The team added obstacles and created a classifier that learned when the simple rope model was reliable but did not attempt to understand the more complex behavior of how the rope interacted with the objects. The team then added recovery steps if the robot encountered a situation, such as when the rope collided with an obstacle, and the classifier determined the simple model was unreliable.
Researchers say their approach drew inspiration from other realms of science and robotics, where simple models are still useful. Using the simple model of a rope, the team developed ways to make sure the object is being used in appropriate situations where the model is reliable, allowing the robot to generalize its knowledge in new situations that have never been encountered.