Researchers from MIT, Amazon Robotics, and the University of British Columbia have developed a new method that allows robots to identify object properties—such as weight and softness—by lifting and shaking them. The technique uses internal sensors instead of cameras or external equipment, making it suitable for environments with limited visibility, like disaster zones.
According to the Massachusetts Institute of Technology, the system relies on joint encoders to detect the robot’s movements and interactions with objects. By simulating these movements, the robot estimates an object’s mass using data from its internal sensors. Chao Liu, an MIT postdoctoral researcher, explained that robots can take advantage of precise measurements from their joints and applied forces, unlike humans.
At the core of the method is a digital simulation, or digital twin, that mirrors the robot's real-world interaction with the object. The simulation predicts how small changes in an object's properties affect the robot's joint positions. According to Peter Yichen Chen, an MIT postdoc and the lead author of the study, an accurate simulation is key to making these predictions.
The technique may be expanded to detect other properties, such as the moment of inertia or the viscosity of fluids. The research is scheduled to be presented at the International Conference on Robotics and Automation.
The method does not replace computer vision but is intended to work alongside it, combining multiple types of sensing for improved performance. The project received support from Amazon and the GIST-CSAIL Research Program and aims to improve robots' ability to learn and adapt in changing environments.
Industry observers note that the approach shows how robots can infer physical properties using only internal data, which could support advancements in autonomous robotics.
