Gestalt Perceptual Calibration for Multi-agent Collaborative Localization

Multi-agent collaborative localization is a challenging problem that requires estimating the global poses of multiple agents in a shared coordinate system based on their sensor data and inter-agent communication. The existing methods, such as map fusion, graph optimization, or consensus algorithms, cannot fully exploit the perceptual information and similarity among different agents, which can improve localization accuracy and consistency. In this paper, we propose a method of gestalt perceptual calibration (GPC) for multi-agent localization, which leverages the power of gestalt psychology to align the coordinate frames of different agents based on their perceptual similarity. We propose a perceptual encoder based on Perceiver model to encode the sensor data of each agent into a latent array. To optimize the relative poses of each pair of agents based on their gestalt similarity, we use a Gumbel-softmax-based relaxation as the gestalt calibrator. Aiming at estimating the global poses of each agent based on the relative poses and other measurements, we also use a maximum a posteriori inference as the pose estimator. The experimental results show that our GPC method can achieve superior localization performance in both accuracy and efficiency compared to existing localization methods. In addition, we further analyze the influence of different parameters of our GPC method on the localization performance.