An Active Robot Object Search Strategy Based on Probabilistic Voronoi Diagram and POMDP

Proactively searching for objects like humans may be a basic requirement for intelligent service robots. The path planning during the searching process can be modeled as a Partially Observable Markov Decision Process (POMDP). In this work, we propose a Probabilistic Voronoi Diagram (PVD) for object search strategy planning on the basis of POMDP. Firstly, an environmental knowledge base is constructed to record the information of objects, and a Bidirectional Encoder Representations from Transformers (BERT) model [1] is trained to encode and decompose the semantic knowledge in the environment. In order to reveal the interrelationships between objects, a Gaussian Mixture Model (GMM) is adopted using the information within the environmental knowledge base. In order to accelerate the searching efficiency, the Generalized Voronoi Diagram (GVD) is introduced to discretize the map and generate the environmental topological map. In order to further establish the spatial correlation between objects, we propose combining the GVD topological map with the GMM to generate the PVD, which can respond to the probability distribution of objects. On the basis of PVD, we further model the object search problem as a POMDP problem by considering the region search cost and the distance traveled cost of the robot in performing solutions. When making observations and updates, the object-to-object relationships in the knowledge base are extracted by the robot to optimize decisions when observing objects related to the target object. Both real-world experimental studies and simulations reveal that our algorithm is very close to human search strategies and outperforms other state-of-the-art algorithms in terms of trajectory length and running time.