A typical assumption in state-of-the-art self-localization models is that an annotated training dataset is available in the target workspace. However, this does not always hold when a robot travels in a general open-world. This study introduces a novel training scheme for open-world distributed robot systems. In our scheme, a robot ("student") can ask the other robots it meets at unfamiliar places ("teachers") for guidance. Specifically, a pseudo-training dataset is reconstructed from the teacher model and thereafter used for continual learning of the student model. Unlike typical knowledge transfer schemes, our scheme introduces only minimal assumptions on the teacher model, such that it can handle various types of open-set teachers, including uncooperative, untrainable (e.g., image retrieval engines), and blackbox teachers (i.e., data privacy). Rather than relying on the availability of private data of teachers as in existing methods, we propose to exploit an assumption that holds universally in self-localization tasks: "The teacher model is a self-localization system" and to reuse the self-localization system of a teacher as a sole accessible communication channel. We particularly focus on designing an excellent student/questioner whose interactions with teachers can yield effective question-and-answer sequences that can be used as pseudo-training datasets for the student self-localization model. When applied to a generic recursive knowledge distillation scenario, our approach exhibited stable and consistent performance improvement.
A typical assumption in state-of-the-art self-localization models is that an annotated training dataset is available for the target workspace. However, this is not necessarily true when a robot travels around the general open world. This work introduces a novel training scheme for open-world distributed robot systems. In our scheme, a robot (``student") can ask the other robots it meets at unfamiliar places (``teachers") for guidance. Specifically, a pseudo-training dataset is reconstructed from the teacher model and then used for continual learning of the student model under domain, class, and vocabulary incremental setup. Unlike typical knowledge transfer schemes, our scheme introduces only minimal assumptions on the teacher model, so that it can handle various types of open-set teachers, including those uncooperative, untrainable (e.g., image retrieval engines), or black-box teachers (i.e., data privacy). In this paper, we investigate a ranking function as an instance of such generic models, using a challenging data-free recursive distillation scenario, where a student once trained can recursively join the next-generation open teacher set.
Cross-view self-localization is a challenging scenario of visual place recognition in which database images are provided from sparse viewpoints. Recently, an approach for synthesizing database images from unseen viewpoints using NeRF (Neural Radiance Fields) technology has emerged with impressive performance. However, synthesized images provided by these techniques are often of lower quality than the original images, and furthermore they significantly increase the storage cost of the database. In this study, we explore a new hybrid scene model that combines the advantages of view-invariant appearance features computed from raw images and view-dependent spatial-semantic features computed from synthesized images. These two types of features are then fused into scene graphs, and compressively learned and recognized by a graph neural network. The effectiveness of the proposed method was verified using a novel cross-view self-localization dataset with many unseen views generated using a photorealistic Habitat simulator.
The emerging ``Floor plan from human trails (PfH)" technique has great potential for improving indoor robot navigation by predicting the traversability of occluded floors. This study presents an innovative approach that replaces first-person-view sensors with a third-person-view monocular camera mounted on the observer robot. This approach can gather measurements from multiple humans, expanding its range of applications. The key idea is to use two types of trackers, SLAM and MOT, to monitor stationary objects and moving humans and assess their interactions. This method achieves stable predictions of traversability even in challenging visual scenarios, such as occlusions, nonlinear perspectives, depth uncertainty, and intersections involving multiple humans. Additionally, we extend map quality metrics to apply to traversability maps, facilitating future research. We validate our proposed method through fusion and comparison with established techniques.
In ground-view object change detection, the recently emerging map-less navigation has great potential as a means of navigating a robot to distantly detected objects and identifying their changing states (appear/disappear/no-change) with high resolution imagery. However, the brute-force naive action strategy of navigating to every distant object requires huge sense/plan/action costs proportional to the number of objects. In this work, we study this new problem of ``Which distant objects should be prioritized for map-less navigation?" and in order to speed up the R{\&}D cycle, propose a highly-simplified approach that is easy to implement and easy to extend. In our approach, a new layer called map-based navigation is added on top of the map-less navigation, which constitutes a hierarchical planner. First, a dataset consisting of $N$ view sequences is acquired by a real robot via map-less navigation. Then, an environment simulator was built to simulate a simple action planning problem: ``Which view sequence should the robot select next?". Then, a solver was built inspired by the analogy to the multi-armed bandit problem: ``Which arm should the player select next?". Finally, the effectiveness of the proposed framework was verified using the semantically non-trivial scenario ``sofa as bookshelf".
The recently emerging research area in robotics, ground view change detection, suffers from its ill-posed-ness because of visual uncertainty combined with complex nonlinear perspective projection. To regularize the ill-posed-ness, the commonly applied supervised learning methods (e.g., CSCD-Net) rely on manually annotated high-quality object-class-specific priors. In this work, we consider general application domains where no manual annotation is available and present a fully self-supervised approach. The present approach adopts the powerful and versatile idea that object changes detected during everyday robot navigation can be reused as additional priors to improve future change detection tasks. Furthermore, a robustified framework is implemented and verified experimentally in a new challenging practical application scenario: ground-view small object change detection.
In this paper, we explore the challenging 1-to-N map matching problem, which exploits a compact description of map data, to improve the scalability of map matching techniques used by various robot vision tasks. We propose a first method explicitly aimed at fast succinct map matching, which consists only of map-matching subtasks. These tasks include offline map matching attempts to find a compact part-based scene model that effectively explains each map using fewer larger parts. The tasks also include an online map matching attempt to efficiently find correspondence between the part-based maps. Our part-based scene modeling approach is unsupervised and uses common pattern discovery (CPD) between the input and known reference maps. This enables a robot to learn a compact map model without human intervention. We also present a practical implementation that uses the state-of-the-art CPD technique of randomized visual phrases (RVP) with a compact bounding box (BB) based part descriptor, which consists of keypoint and descriptor BBs. The results of our challenging map-matching experiments, which use a publicly available radish dataset, show that the proposed approach achieves successful map matching with significant speedup and a compact description of map data that is tens of times more compact. Although this paper focuses on the standard 2D point-set map and the BB-based part representation, we believe our approach is sufficiently general to be applicable to a broad range of map formats, such as the 3D point cloud map, as well as to general bounding volumes and other compact part representations.
Semantic localization, i.e., robot self-localization with semantic image modality, is critical in recently emerging embodied AI applications such as point-goal navigation, object-goal navigation and vision language navigation. However, most existing works on semantic localization focus on passive vision tasks without viewpoint planning, or rely on additional rich modalities (e.g., depth measurements). Thus, the problem is largely unsolved. In this work, we explore a lightweight, entirely CPU-based, domain-adaptive semantic localization framework, called graph neural localizer.Our approach is inspired by two recently emerging technologies: (1) Scene graph, which combines the viewpoint- and appearance- invariance of local and global features; (2) Graph neural network, which enables direct learning/recognition of graph data (i.e., non-vector data). Specifically, a graph convolutional neural network is first trained as a scene graph classifier for passive vision, and then its knowledge is transferred to a reinforcement-learning planner for active vision. Experiments on two scenarios, self-supervised learning and unsupervised domain adaptation, using a photo-realistic Habitat simulator validate the effectiveness of the proposed method.
Visual place classification from a first-person-view monocular RGB image is a fundamental problem in long-term robot navigation. A difficulty arises from the fact that RGB image classifiers are often vulnerable to spatial and appearance changes and degrade due to domain shifts, such as seasonal, weather, and lighting differences. To address this issue, multi-sensor fusion approaches combining RGB and depth (D) (e.g., LIDAR, radar, stereo) have gained popularity in recent years. Inspired by these efforts in multimodal RGB-D fusion, we explore the use of pseudo-depth measurements from recently-developed techniques of ``domain invariant" monocular depth estimation as an additional pseudo depth modality, by reformulating the single-modal RGB image classification task as a pseudo multi-modal RGB-D classification problem. Specifically, a practical, fully self-supervised framework for training, appropriately processing, fusing, and classifying these two modalities, RGB and pseudo-D, is described. Experiments on challenging cross-domain scenarios using public NCLT datasets validate effectiveness of the proposed framework.