University of Washington researchers have developed a groundbreaking method called THOR (Topology-based Hybrid Object Recognition) to teach robots the skill of identifying objects, even when they are partially obscured or in cluttered spaces. This innovation addresses a significant challenge faced by robots in warehouses and households, where they often struggle to accurately identify and pick up objects due to the lack of "object unity" – the ability to recognize objects even when not fully visible.
THOR allows robots to create a 3D representation of objects based on their shapes and then uses topology, a branch of mathematics, to assign each object to a "most likely" object class. Unlike previous approaches that rely on machine learning models trained with images of cluttered rooms, THOR only requires images of individual objects, making it more efficient and versatile. Moreover, THOR does not require specialized sensors or processors, making it cost-effective and easily implementable with commodity cameras.
Lead author Ekta Samani, a UW doctoral student, developed THOR to mimic human perception of partially visible objects, enabling robots to accurately identify objects in real-time, even in diverse environments with varying backgrounds, lighting conditions, and degrees of clutter. The method outperforms existing 3D shape-based recognition methods by providing more detailed object representations.
THOR has wide-ranging applications and can be integrated into any indoor service robot operating in various settings such as homes, offices, stores, warehouses, and manufacturing plants. While THOR excels in identifying all types of objects in cluttered spaces, it performs particularly well with kitchen-style objects like mugs and pitchers due to their distinctive but regular shapes.
The researchers are exploring additional enhancements to THOR, including incorporating color, texture, and text labels into object recognition, addressing squishy or damaged objects, and enabling robots to navigate around obstructed objects for better visibility. Additionally, they are working on enabling robots to categorize unfamiliar objects and seek human assistance for correct identification.
THOR represents a significant advancement in robot object recognition and has the potential to revolutionize robotics applications in various industries, making robots more adept at navigating complex environments and performing tasks with precision and efficiency.
