Teaching Robots Object Unity: The Puzzle of Identifying Objects in Cluttered Spaces

Sunday - February 11 2024, 09:01 UTC - 9 months ago

Researchers at the University of Washington have developed a method called THOR, which allows robots to identify objects in cluttered spaces. THOR outperformed current models and is trainable, making it adaptable to new environments. This method uses a deep-learning approach and an information-theoretic approach to select optimal angles for image collection.

As robots become more prevalent in our daily lives, researchers and engineers are constantly working to improve their capabilities. One particular challenge that robots face is identifying objects in cluttered spaces. For humans, this task seems relatively easy - we can still recognize an object even if we can't see all of it. But for robots, this skill is not as intuitive.

Researchers at the University of Washington have developed a solution to this problem, called THOR. This method allows robots to identify objects even in cluttered spaces, and has shown to outperform current state-of-the-art models.

The team used a low-cost robot, equipped with vision sensors, to test their method. The robot, similar to the human brain's visual cortex, processes images to detect objects, estimate their orientations, and identify what they are. But in cluttered spaces, this can be a difficult task for robots. There are a large number of objects of varying shapes and sizes, making it hard to distinguish between different object types. Additionally, objects that are too close together can obstruct the robot's view, making it difficult to recognize objects.

While this may seem like a straightforward problem to solve, previous attempts have not been as successful as THOR. So what makes THOR different and better? .

First, THOR uses a deep-learning approach to identify objects based on their textures and shapes. This is similar to how our brains work - objects have unique textures that help us identify them. By incorporating this into THOR's algorithm, the robot is better able to identify objects in cluttered spaces. Additionally, THOR uses an information-theoretic approach to select the best vantage points for image collection. This means that the robot can choose the most optimal angle to take images from, maximizing its chances of correctly identifying the objects.

And perhaps most impressively, THOR is trainable. This means that after initial training, the robot can continue to learn and adapt to new objects and environments. This flexibility is crucial for robots as they are often used in different spaces and settings.

Senior author Ashis Banerjee, UW associate professor, explained that THOR is the brainchild of lead author Ekta Samani, a UW doctoral student. With THOR's success, the potential for robots to navigate and operate in cluttered spaces has greatly increased, opening up new possibilities for their use in daily lives.