Object Shape Recognition Using Kriging-Based Spatial Estimation and Reinforcement Learning
Abstract:
Object shape recognition is a critical component of robotic perception, enabling autonomous systems to understand, interact with, and manipulate objects in their environment. In real-world applications such as search and rescue, warehouse logistics, and planetary exploration, robots often encounter unstructured environments where sensor data is sparse, noisy, or occluded—conditions under which traditional vision-based recognition methods are prone to failure. To overcome these limitations, this paper proposes a hybrid framework that integrates Kriging-based spatial estimation with reinforcement learning (RL) to enable robust and adaptive shape inference. Kriging, a probabilistic geostatistical interpolation technique, is employed to reconstruct continuous surface representations from sparse sensor measurements by exploiting spatial correlations modeled via variograms. This approach yields not only accurate surface predictions but also spatial uncertainty maps, which are used to guide the sensing process. Complementing this model-based estimation, an RL agent is trained to actively explore the environment by selecting sensing actions that maximize information gain while minimizing redundancy. The agent’s policy is optimized to identify high-uncertainty regions and prioritize them during exploration, thereby improving both the efficiency and accuracy of shape reconstruction. The proposed method is evaluated in a series of simulated experiments involving synthetic 3D objects with diverse geometries, occlusion patterns, and varying levels of sensor noise. Results demonstrate that the Kriging-RL framework achieves over 92% shape reconstruction accuracy, reduces exploration time by up to 35% compared to passive methods, and generalizes well to unseen and partially observed shapes. Ablation studies further highlight the contribution of the RL component in enhancing exploration strategies, while robustness tests confirm the system’s stability under noisy conditions. This work establishes a promising foundation for integrating statistical modeling and adaptive control in robotic perception systems, enabling efficient and resilient shape recognition in complex and uncertain environments.
