Spatial Planning For Placement Command Understanding

Object placement in virtual and real worlds is an important task for autonomous agents and applications. Interacting with agents using natural language commands presents an intuitive alternative to graphical and other operator control interfaces. However, understanding and interpreting language for placement actions in 3D continuous spaces is computationally hard. In this paper, we extend our previous work on resolving underspecified linguistic commands for object placement using a spatial planning approach called object placement using ordered application and simulation of constraints (OPOCS). This heuristic approach represents and utilizes practical real-world knowledge of objects an their interactions. It does this with simulation of human activities to accurately place and orient objects. We evaluated OPOCS and compared it with a naïve algorithm for understanding 10 representative spatial terms and relations in 4 different 3D office worlds. Our findings show that OPOCS significantly outperforms a naïve algorithm.