Among the multitude of scenarios in which autonomous robots are intended to operate, natural environments present the most significant challenges in the context of traversability estimation compared to structured settings. To address these complexities without compromising urban performance, we propose TE-NeXt (Traversability Estimation Convolutional Network), a customized 3D sparse convolutional architecture tailored for unstructured terrain. This network is a customized and efficient architecture for traversability estimation from sparse LiDAR (Light Detection and Ranging) point clouds based on a encoder–decoder topology that includes several modifications regarding: (i) the input features; (ii) the structure of encoder–decoder resolution levels; and (iii) the constitution of the 3D (three-dimensional) convolutional block. Thus, the experimental results demonstrate superior performance in unstructured terrain (82% F1 score on Rellis-3D), high robustness in urban environments (SemanticKITTI), and strong generalization capabilities in mixed environments (SemanticUSL). Finally, we present a fully autonomous navigation framework utilizing this method and release the source code to ensure reproducibility.

This paper presents CrossPlace, an innovative method for cross-modal place recognition between heterogeneous sensor modalities, particularly between fisheye cameras and LiDAR. Place recognition is the fundamental capability of mobile robots to determine their most likely location within a database, based on sensory input queries. In cross-modal place recognition, the goal is to localize using a different sensor from the one originally used to construct the database. The core contribution of this paper is a unified feature space that integrates intensity, depth and semantic information. Both the database entries and the queries are obtained by embedding sensor readings through the same CrossPlace model, ensuring a consistent representation across modalities. Consequently, a database constructed from LiDAR can be queried with fisheye images, and vice versa, using a single shared architecture. Furthermore, a comprehensive data transformation and preprocessing pipeline is presented. Specifically, CrossPlace is constituted by three independent branches, each one for processing intensity, depth and semantic information. Each branch consists of a CosPlace model for image embedding with shared weights across sensor modalities. Late fusion through concatenation of the intensity, depth and semantic embeddings provides optimal global performance. We conduct an exhaustive evaluation on the KITTI-360 dataset, where CrossPlace surpasses state-of-the-art techniques across all metrics, establishing a new standard for cross-modal place recognition in urban and highway environments. The results demonstrate the effectiveness of our unified approach for place recognition across different sensor modalities while maintaining a robust performance under various operating environments.

The location of a robot is a key aspect in the field of mobile robotics. This problem is particularly complex when the initial pose of the robot is unknown. In order to find a solution, it is necessary to perform a global localization. In this paper, we propose a method that addresses this problem using a coarse‐to‐fine solution. The coarse localization relies on a probabilistic approach of the Monte Carlo localization (MCL) method, with the contribution of a robust deep learning model, the MinkUNeXt neural network, to produce a robust description of point clouds of a 3D LiDAR within the observation model. The MCL method has been approached from a topological perspective, considering that the particles are initialized on the map positions where LiDAR scans have been previously captured. For fine localization, global point cloud registration has been implemented. MinkUNeXt aids this by exploiting the outputs of its intermediate layers to produce deep local features for each point in a scan. These features facilitate precise alignment between the current sensor observation (query) and one of the point clouds on the map. The proposed MCL method incorporating deep local features for fine localization is termed MCL‐DLF. Alternatively, a classical ICP method has been implemented for this precise localization aiming at comparison purposes. This method is termed as MCL‐ICP. In order to validate the performance of the MCL‐DLF method, it has been tested on publicly available datasets such as the NCLT dataset, which provides seasonal large‐scale environments. In addition, tests have been also performed with our own data (UMH) that also include seasonal variations on large indoor/outdoor scenarios. The results, which were compared with established state‐of‐the‐art methodologies, demonstrate that the MCL‐DLF method obtains an accurate estimate of the robot localization in dynamic environments despite changes in environmental conditions. For reproducibility purposes, the code is publicly available.

Omnidirectional cameras are a suitable and cost-effective choice for Visual Place Recognition (VPR), as they provide comprehensive information from the scene regardless of the robot orientation. However, vision sensors are vulnerable to environmental appearance changes (e.g., illumination, weather, season or moving objects). While multi-modal sensing approaches can overcome these challenges, they introduce significant cost and system complexity. This paper introduces PDPR (Panoramic-Depth Place Recognition), a novel fusion framework that enhances the robustness of VPR methods by integrating visual data with geometric features derived from monocular depth estimation techniques, while using a single-camera setup. In the ablation study, both early and late fusion strategies are evaluated to optimally combine appearance-based and depth-derived features. The extensive evaluation on challenging, indoor and outdoor datasets demonstrates that PDPR consistently boosts retrieval performance across multiple state-of-the-art VPR models. Furthermore, this improvement is achieved without requiring any fine tuning, allowing our method to function as a pluggable module for pretrained models. Consequently, this work presents a powerful, practical and low-cost solution for robust VPR, with high potential to scale as monocular depth estimation and VPR models continue to improve. The project website can be found at https://marcosalfaro.github.io/projects-PDPR/ . • Monocular depth estimation is used to enhance place recognition. • A thorough evaluation of preprocessing techniques to enhance the depth maps. • Fusion techniques are designed to leverage visual and geometric data. • A model-agnostic approach that improves the performance even with no fine tuning. • A robust method across different scenarios and lighting conditions.

Redundant manipulators offer several advantages, including improved manipulability, singularity avoidance, and obstacle evasion. However, kinematic redundancy also introduces additional challenges, such as the need to solve an underdetermined inverse kinematic problem to control the manipulator. This paper introduces a novel approach for motion planning of redundant manipulators, based on the exploration of feasibility maps. The proposed method is an extension of the RRT algorithm, modified to explore the redundant space in order to find a suboptimal feasible path in the joint space, sacrificing optimality for scalability to higher degrees of redundancy. The method is able to follow a given task trajectory while considering other constraints, such as joint limits, self-collisions, and obstacles.

Feature matching is a key technique for a wide variety of computer vision and image processing applications such as visual localization. It permits finding correspondences of significant points within the environment that eventually determine the localization of a mobile agent. In this context, this work evaluates an Adaptive Probability-Oriented Feature Matching (APOFM) method that dynamically models the visual knowledge of the environment in terms of the probability of existence of features. Several improvements are proposed to achieve a more robust matching in a visual odometry framework: a study on the classification of the matching candidates, enhanced by a nearest neighbour search policy; a dynamic weighted matching that exploits the probability of feature existence in order to tune the matching thresholds; and an automatic false positive detector. Additionally, a comparison of performance is carried out, considering a publicly available dataset composed of two kinds of wide field-of-view images: catadioptric and fisheye. Overall, the results validate the appropriateness of these contributions, which outperform other well-recognized implementations within this framework, such as the standard visual odometry, a visual odometry method based on RANSAC, as well as the basic APOFM. The analysis shows that fisheye images provide more visual information of the scene, with more feature candidates. Contrarily, omnidirectional images produce fewer feature candidates, but with higher ratios of feature acceptance. Finally, it is concluded that improved precision is obtained when the location problem is solved by this method.

This paper presents the workspace analysis of a redundant serial-parallel robot. Due to the complexity of the robot, the complex constraints (joint limits and no-interference between the legs of the robot), and the globally serial structure of the robot, a discretization method based on the forward kinematics would be most appropriate to compute the workspace. However, this widely used method can only obtain the external boundaries of the workspace, missing the internal barriers that hinder the motion of the robot, which may exist inside the boundaries. To avoid missing these barriers, we use a discretization method that uses the solution of the inverse kinematic problem of the robot. By studying the feasibility of attaining a desired position and orientation by the different branches of the solution to the inverse kinematics, the proposed discretization method is able to obtain both the external boundaries and the internal barriers of the workspace. Some examples are presented to show the importance of these internal barriers in the motions of the robot inside the workspace.

This paper presents a novel climbing robot that explores 3-D truss structures for maintenance and inspection tasks. The robot is biped and has a hybrid serial-parallel architecture since each leg is composed of two parallel mechanisms connected in series. First, the forward kinematic problem of the complete robot is solved, obtaining the relative position and orientation between the feet in terms of the ten joint coordinates of the robot. The inverse kinematics is more complex due to the redundancy of the robot. Hence, a simplified inverse kinematic problem that assumes planar and symmetric movements is analyzed. Then, a tool to simulate the kinematics of the robot is presented, and it is used to demonstrate that the robot can completely explore 3-D structures, even when some movements are restricted to be planar and symmetric.

This work presents a purely visual topologic odometry system for robot navigation. Our system is based on a Multi-Scale analysis that allows us to estimate the relative displacement between consecutive omnidirectional images. This analysis uses global appearance techniques to describe the scenes. The visual odometry system also makes use of global appearance descriptors of panoramic images to estimate the phase lag between consecutive images and to detect loop closures. When a previous mapped area is recognized during the navigation, the system re-estimates the pose of the scenes included in the map, reducing the error of the path. The algorithm is validated using our own database captured in an indoor environment under real dynamic conditions. The results demonstrate that our system permits estimating the path followed by the robot with accuracy comparing to the real route.

In this paper we compare two methods to carry out topological mapping using only visual information captured by a robot. This map should contain enough information so that the robot can estimate its position and orientation and redundant information should be removed to get an acceptable computational cost during the localization process. Apart from this, it is also important to know the topology of the map created since it will make possible a high-level planification of the path to move to the target points. We propose to build this topological map only using the panoramic images taken by an omnidirectional vision system and using appearance-based methods. We have carried out an exhaustive experimentation to study the validity of the proposed methods and to perform an objective comparison between them. Also, we have tested the processing time to create the topological map.

In a multi-robot system, in which each of the robots constructs its own local map, it is necessary to perform the fusion of these maps into a global one. This task is normally performed in two different steps: by aligning the maps and then merging the data. This paper focusses on the first step: Map Alignment, which consists in obtaining the transformation between the local maps built independently. In this way, these local maps will have a common reference frame. In this paper, a collection of algorithms for solving the map alignment are analyzed under different conditions of noise in the data and intersection between local maps. This study is performed in a visual SLAM context, in which the robots construct landmark-based maps. The landmarks consist in 3D points captured from the environment and characterized by a visual descriptor.

This paper presents an approach to vision-based simultaneous localization and mapping (SLAM). Our approach uses the scale invariant feature transform (SIFT) as features and applies a rejection technique to concentrate on a reduced set of distinguishable, stable features. We track detected SIFT features over consecutive frames obtained by a stereo camera and select only those features that appear to be stable from different views. Whenever a feature is selected, we compute a representative feature given the previous observations. This approach is applied within a Rao-Blackwellized particle filter to make the data association easier and furthermore to reduce the number of landmarks that need to be maintained in the map. Our system has been implemented and tested on data gathered with a mobile robot in a typical office environment. Experiments presented in this paper demonstrate that our method improves the data association and in this way leads to more accurate maps