Improved and scalable online learning of spatial concepts and language models with mapping

Abstract

We propose a novel online learning algorithm, called SpCoSLAM 2.0, for spatial concepts and lexical acquisition with high accuracy and scalability. Previously, we proposed SpCoSLAM as an online learning algorithm based on unsupervised Bayesian probabilistic model that integrates multimodal place categorization, lexical acquisition, and SLAM. However, our original algorithm had limited estimation accuracy owing to the influence of the early stages of learning, and increased computational complexity with added training data. Therefore, we introduce techniques such as fixed-lag rejuvenation to reduce the calculation time while maintaining an accuracy higher than that of the original algorithm. The results show that, in terms of estimation accuracy, the proposed algorithm exceeds the original algorithm and is comparable to batch learning. In addition, the calculation time of the proposed algorithm does not depend on the amount of training data and becomes constant for each step of the scalable algorithm. Our approach will contribute to the realization of long-term spatial language interactions between humans and robots.

Spatial concept-based navigation with human speech instructions via probabilistic inference on Bayesian generative model

“…We introduce the outline of SpCoSLAM [5] as an unsupervised learning method for the acquisition of spatial knowledge to use in a navigational task. Our study is specific to the validation of navigation tasks, and more detailed information regarding learning algorithms has been presented in previous reports [5,6].…”

“…Further, this method can estimate an appropriate number of clusters of spatial concepts and position distributions depending on the data using the so-called online Chinese restaurant process [25]. Additionally, to enable long-term learning for a real robot with limited calculation resources, a scalable and improved online learning algorithm was developed [6].…”

“…The details of the properties of the robot are described in the dataset. Further, we used the global parameters G (map, parameters of spatial concept, and language model) of the maximum likelihood particle at step 50 learned in SpCoSLAM 2.0 [6]. We used the navigation and costmap_2d package in the ROS to obtain a global cost map.…”

“…As a semantic mapping method including place categorization, a Bayesian generative model named SpCoSLAM that acquired spatial concepts, i.e. categorical knowledge of places from multimodal information through unsupervised learning, has been proposed [5,6]. In this paper, we focus on a method by which a robot utilized spatial concepts learned from its own experience for navigational tasks.…”

“…Overview of navigational tasks with human speech instruction. Top: Learning phase for multimodal place categorization based on unsupervised learning in SpCoSLAM [5,6]. (i) The user teaches the robot the place name by an uttered sentence, while moving together.…”