
In artificial intelligence (AI), commonsense reasoning refers to the ability to make assumptions about the characteristics and nature of everyday situations, similar to how humans perceive and interpret them. For example, it involves evaluating physical objects, their properties, and their interactions with the surroundings, which is the focus of this investigation. The authors present a hybrid system that combines commonsense reasoning, deep learning, and continuous learning to solve problems in dynamic and uncertain environments. In particular, they focus on an example of a robot that reduces clutter by estimating object occlusion and stability in scenes.
The system design is practical, integrating reasoning through answer set programming (ASP) and deep learning using convolutional neural networks (CNNs). The ASP module handles incomplete knowledge through rules and axioms, while CNNs address complex tasks by focusing on areas where reasoning fails. The system also learns new information gradually, adapting to changing situations. It combines qualitative spatial representations with learned metric spatial representations to understand relationships such as “above,” “left_of,” or “near.” A human-like forgetting mechanism helps manage and update learned information.
The system was tested with both simulated and real-world data (the paper includes images of improvised scenes with objects made from cornflake boxes), demonstrating improved accuracy and efficiency compared to deep learning alone. Key results include better performance in occlusion and stability tasks, reduced reliance on large labeled datasets, and the ability to refine knowledge over time.
While the study highlights the benefits of combining reasoning and learning, it focuses on a specific robot task and does not assess the system’s scalability to more complex scenarios. Future research should explore scalability, particularly given the computational demands of ASP. Although COVID-19 restrictions prevented real-world robot testing, the system shows strong potential for assistive robotics. Additionally, the authors have implemented an elementary task planning system; however, the area of task and motion planning (TAMP) research has been substantially explored, and other publications [1,2,3] might offer additional ideas and insights. Overall, the paper is interesting, the research is well executed, and I sincerely hope future work includes real-world experiments with physical robots.