Mass Prediction and Analysis of an Object’s Mass from Its Image Using Deep Learning

The world is full of objects with different properties, such as volume, density, mass, material, etc. These properties are crucial for meaningful interaction with these objects. However, none of these properties is explicitly known. As humans rely on perception to interact with the world, robots need to understand the properties of objects to function effectively. This includes inferring properties that may not be explicitly stated, such as an object’s mass. A single robot might not be sufficient to execute complex tasks in dynamic environments. In such cases, an agent can assemble a team of robots at runtime to complete the task. The agent must estimate the object’s mass to make this decision effectively. For example, if a team of robots needs to move a heavy object, accurately estimating its mass is essential. This information allows the agent to determine the appropriate team size, ensuring enough collective strength to complete the task successfully. In this work, we describe techniques that enable robots to learn to fill these knowledge gaps (missing properties) before attempting to interact with these objects. The salient features of our work lie in our observation that the properties of an object are not entirely independent of each other, and learning to predict/infer one can give clues regarding predicting/inferring the other. Our evaluation indicates that the proposed solutions outperform the existing state-of-the-art method. We also simulated estimating the object’s properties in the ROS environment.