Effective Visual Localization (VL) requires a map of the environment that combines compactness for efficient scalability with robustness against visual appearance changes and metric precision. Through low-dimensional image embeddings, Visual Place Recognition (VPR) is able to successfully meet the first two requirements, but its low metric accuracy makes it less suitable than standard VL approaches based on local features or neural representations. This limitation can be overcome by integrating VPR with the accurate local trajectory estimates produced by feed-forward neural 3D geometry (FF3D) models. In this paper, we address sequential appearance-based localization through a topometric framework that iteratively combines probabilistic VPR with FF3D metric pose estimation in controlled image sets. Our approach proposes an automatic offline mapping tool that models the topometric pose-appearance interaction in the different parts of the scene. This map is later employed by an online particle filter that estimates the pose from odometry and belief over places for FF3D inference, successfully incorporating neural metric estimation into probabilistic appearance-based localization. We extensively evaluate the framework on three known benchmarks, demonstrating substantial improvements over existing appearance-based methods. The modularity of our approach allows the descriptor extractor and FF3D model to remain interchangeable, and a focused analysis further shows that sequential belief can mitigate severe failures under perceptual aliasing.

Robotic applications increasingly rely on distributed computational infrastructures that combine embedded devices, edge servers, and cloud resources. This evolution, together with the collaborative nature of robotics projects, has made the development, integration, deployment, and long-term operation of robotic systems significantly more complex. In practice, multi-user robotics software teams face persistent challenges related to dependency isolation, compatibility, reproducibility, efficient sharing of specialized hardware, and deployment across heterogeneous environments. In this paper, we present CSAR (Containerized System Architecture for Robotics), a container-centric architectural framework designed specifically for robotics teams and the edge-cloud continuum. CSAR combines LXC/LXD-based system containerization, ROS 2/DDS-based communication, and a three-layer edge infrastructure to organize computation into hardware-affine, persistent execution environments that remain decoupled from the volatility of experimental workloads. Through its Infrastructure Core, Platform and Multi-User Orchestration, and Compute and Acceleration layers, CSAR provides strong isolation, controlled resource sharing, and topology-aware networking for distributed robotic applications. To demonstrate its validity, we describe a real deployment of CSAR in an academic robotics laboratory and evaluate it through representative use cases involving edge-offloaded 3D SLAM and GPU-accelerated semantic mapping. The results indicate that CSAR simplifies software integration, improves the utilization of shared computational resources, and facilitates safe prototyping, as well as reproducible and collaborative experimentation in robotics teams. The implementation described in this paper, including deployment templates, configuration files, and documentation, is available at https://github.com/goyoambrosio/CSAR.

Recent advances in Generative Artificial Intelligence (GenAI), particularly Large Language Models (LLMs), are driving robotic architectures toward agent-based high-level orchestration, in which natural-language instructions can be translated into context-aware action sequences. While the integration of these agents and robotic capabilities is increasingly converging toward standardization through the Model Context Protocol (MCP), the upper Human-Robot Interaction (HRI) layer remains fragmented by proprietary, ad hoc interfaces that hinder real-time human-in-the-loop collaboration. To address this fragmentation, this paper proposes the adoption of the Agent-Client Protocol (ACP) – a communication standard originally introduced for coding agents in software engineering – as a unified communication contract for the HRI layer in agent-based robotic systems. By combining ACP at the interface-agent link and MCP at the agent-execution link, we formulate a fully decoupled three-layer architecture that separates human interaction, deliberative orchestration, and physical execution. This topology removes rigid architectural dependencies, enabling heterogeneous user interfaces to connect to the same robotic system and allowing the underlying robotic platform to be replaced without requiring client-specific integration changes. Moreover, it provides native support for collaborative HRI capabilities such as real-time observability, explicit human authorization, and immediate task interruption. We experimentally evaluate the proposed architecture on a physical mobile robot, demonstrating interoperability across three heterogeneous user interfaces and validating real-time human-in-the-loop workflows with negligible latency overhead.

Robots in human-centered environments require accurate scene understanding to perform high-level tasks effectively. This understanding can be achieved through instance-aware semantic mapping, which involves reconstructing elements at the level of individual instances. Neural networks, the de facto solution for scene understanding, still face limitations such as overconfident incorrect predictions with out-of-distribution objects or generating inaccurate masks. Placing excessive reliance on these predictions makes the reconstruction susceptible to errors, reducing the robustness of the resulting maps and hampering robot operation. In this work, we propose Voxeland, a probabilistic framework for incrementally building instance-aware semantic maps. Inspired by the Theory of Evidence, Voxeland treats neural network predictions as subjective opinions regarding map instances at both geometric and semantic levels. These opinions are aggregated over time to form evidence, and are formalized through a probabilistic model. This enables us to quantify uncertainty in the reconstruction process, facilitating the identification of map areas requiring improvement (e.g. reobservation or reclassification). As a possible strategy to exploit this uncertainty quantification, we incorporate a Large Vision-Language Model (LVLM) to perform semantic level disambiguation for instances with high uncertainty. Results from the standard benchmarking on the publicly available SceneNN dataset demonstrate that Voxeland outperforms state-of-the-art methods, highlighting the benefits of incorporating and leveraging both instance- and semantic-level uncertainties to enhance reconstruction robustness. This is further validated through qualitative and quantitative experiments conducted on the real-world ScanNet dataset.

An accurate characterization of indoor airflow is essential for assessing air quality, understanding airborne contaminant transport, and evaluating ventilation performance in buildings. In practice, however, obtaining spatially resolved airflow information remains challenging: high-fidelity CFD simulations are computationally intensive, while interpolation of sparse sensor data fails to enforce the underlying physical behavior of the flow.This paper presents a physics-informed mapping approach based on Gaussian Markov Random Fields (GMRF) to reconstruct two-dimensional airflow fields from sparse internal observations. The method embeds simplified physical constraints derived from the Navier–Stokes equations—including mass conservation, advection, and viscous diffusion—into the estimation process, ensuring physically consistent flow reconstruction at a fraction of the computational cost of CFD.Experimental results demonstrate that the proposed approach accurately recovers airflow structures in a range of indoor environments and remains robust under sparse and irregular sampling conditions. Furthermore, it enables the identification of poorly ventilated (stagnant) regions and the characterization of airflow pathways, providing a practical tool for ventilation assessment and real-time indoor air quality analysis.

Large Language Models (LLMs) provide cognitive capabilities that enable robots to interpret and reason about their workspace, especially when paired with semantically rich representations like semantic maps. However, these models are prone to generating inaccurate or invented responses, known as hallucinations, that can produce an erratic robotic operation. This can be addressed by employing agentic workflows, structured processes that guide and refine the model’s output to improve response quality. This work formally defines and qualitatively analyzes the impact of three agentic workflows (LLM Ensemble, Self-Reflection, and Multi-Agent Reflection) on enhancing the reasoning capabilities of an LLM guiding a robotic system to perform object-centered planning. In this context, the LLM is provided with a pre-built semantic map of the environment and a query, to which it must respond by determining the most relevant objects for the query. This response can be used in a multitude of downstream tasks. Extensive experiments were carried out employing state-of-the-art LLMs and semantic maps generated from the widely-used datasets ScanNet and SceneNN. Results show that agentic workflows significantly enhance object retrieval performance, especially in scenarios requiring complex reasoning, with improvements averaging up to 10% over the baseline.

A mobile robot's ability to perform complex tasks depends critically on its understanding of the environment in which it operates. This understanding-often represented as a map-must incorporate both semantic information about objects (identity, functionality, ...) and a high-level grasp of spatial organization. This latter capability, referred to as place segmentation and categorization, is essential for reasoning in diverse and dynamic spaces. However, conventional semantic mapping methods typically segment environments into pre-defined rooms and rely on closed-set labels, limiting adaptability and expressiveness in real-world, multi-functional settings. In this paper, we propose a novel pipeline for indoor place segmentation and categorization that leverages rich object semantics to generate more flexible and meaningful spatial representations. Rather than adhering to rigid architectural boundaries, our approach defines places as clusters of functionally coherent objects. These clusters are formed using descriptors that integrate geometric location with semantic information, derived from context-aware functional descriptions of objects generated by Large Language Models (LLMs) and encoded as sentence embeddings. We apply density-based clustering to identify semantically meaningful places and use LLMs to assign each cluster open-set natural-language tags and detailed functional descriptions, resulting in enriched and interpretable maps. Experimental results on the ScanNet and SceneNN datasets demonstrate that our method significantly improves segmentation consistency over geometric and basic semantic baselines while producing coherent and informative place categorizations that better reflect the functional layout of indoor environments.

La navegaci´on segura es un aspecto cr´ıtico para los robots sociales que se desplazan en entornos concurridos. Adem´as, estosrobots deben respetar los espacios personales para evitar generar desconfianza y, al mismo tiempo, interactuar de forma efectiva ysocialmente aceptable. Este trabajo presenta un sistema multimodal que combina t´ecnicas de visi´on por computador, planificaci´onde trayectorias socialmente aceptables y estrategias de comunicaci´on proactiva con las personas cercanas al robot. El sistema hasido validado en escenarios controlados, mostrando resultados preliminares prometedores en t´erminos de robustez y fluidez social.

Semantic scene understanding allows a robotic agent to reason about problems in complex ways, using information from multiple and varied sensors to make deductions about a particular matter. As a result, this form of intelligent robotics is capable of performing more complex tasks and achieving more precise results than simpler approaches based on single data sources. However, these improved capabilities come at the cost of higher complexity, both computational and in terms of design. Due to the increased design complexity, formal approaches for exploiting semantic understanding become necessary. We present here a probabilistic formulation for integrating semantic knowledge into the process of gas source localization (GSL). The problem of GSL poses many unsolved challenges, and proposed solutions need to contend with the constraining limitations of sensing hardware. By exploiting semantic scene understanding, we can leverage other sources of information, such as vision, to improve the estimation of the source location. We show how our formulation can be applied to pre-existing GSL algorithms and the effect that including semantic data has on the produced estimations of the location of the source.

This paper presents a Top-view Transformation Model (TTM) for a vision-based autonomous navigation of an agricultural mobile robot. The TTM transforms images captured by an onboard camera into a virtual Top-view, eliminating perspective distortions such as the vanishing point effect and ensuring uniform pixel distribution. The transformed images are analyzed to ensure an autonomous navigation of the robot between crop rows. The navigation method involves real-time estimation of the robot's position relative to crop rows and the control low is derived from the estimated robot's heading and lateral offset for steering the robot along the crop rows. A simulated scenario has been generated in Gazebo in order to implement the developed approach using the Robot Operating System (ROS), while an evaluation on a real agricultural mobile robot has also been performed. The experimental results demonstrate the feasibility of the TTM approach and its implementation for autonomous navigation, reaching good performance.

The dispersion of volatile substances in the air is a complex phenomenon that is relevant to many practical problems. The study of this issue and the development of techniques to deal with it greatly benefit from the use of accurate simulation tools. However, simulation with numerical CFD models is not always applicable, particularly in cases where computational cost is a concern or those that require integration with external software such as robotic platforms. In this article, we present Gaden-RT, which is based on the Gaden simulator, a widely employed tool in the field of robotics research, which uses the filament model to simulate the dispersion of gas in complex environments. We present multiple new features and improvements, including a modified workflow, which allows for faster-than-real-time simulation and online interaction. Thus, Gaden-RT introduces novel simulation capabilities to the state of the art by allowing for the modeling of the airflow disturbances caused by moving robotic agents. The software is made available as an open-source project through an on-line repository.11 https://github.com/MAPIRlab/gaden_core.

Los mapas semánticos son representaciones del entorno que incluyen información sobre la geometría de la escena y la clasificación en categorías de los objetos presentes.En este trabajo se proponen técnicas para el refinamiento de mapas semánticos mediante la desambiguación de los objetos con alta incertidumbre en su clasificación semántica. Concretamente, nuestra propuesta consiste en la identificación sistemática de aquellas instancias que requieren un proceso adicional de desambiguación, y el empleo de un modelo de visión-lenguaje (LVLM) para llevar a cabo este proceso.La implementación de nuestra propuesta se basa en Voxeland, un marco de trabajo que construye dichos mapas siguiendo un enfoque probabilístico, permitiendo así cuantificar la incertidumbre en las clasificaciones de objetos.Las pruebas realizadas sobre el conjunto de datos de SceneNN validan la efectividad del método, mejorando la clasificación de los objetos y reduciendo la incertidumbre de los mapas semánticos.

Large Language Models (LLMs) provide cognitive capabilities that enable robots to interpret and reason about their workspace, especially when paired with semantically rich representations like semantic maps. However, these models are prone to generating inaccurate or invented responses, known as hallucinations, that can produce an erratic robotic operation. This can be addressed by employing agentic workflows, structured processes that guide and refine the model’s output to improve response quality. This work formally defines and qualitatively analyzes the impact of three agentic workflows (LLM Ensemble, Self-Reflection, and Multi-Agent Reflection) on enhancing the reasoning capabilities of an LLM guiding a robotic system to perform object-centered planning. In this context, the LLM is provided with a pre-built semantic map of the environment and a query, to which it must respond by determining the most relevant objects for the query. This response can be used in a multitude of downstream tasks. Extensive experiments were carried out employing state-of-the-art LLMs and semantic maps generated from the widely-used datasets ScanNet and SceneNN. The results show that agentic workflows significantly enhance object retrieval performance, especially in scenarios requiring complex reasoning, with improvements averaging up to 10% over the baseline.

The early detection of mild cognitive impairment, a condition of increasing impact in our aging society, is a challenging task with no established answer. One promising solution is the deployment of robotic systems and ambient assisted living technology in the houses of older adults for monitoring and assistance. In this work, we address and discuss a qualitative analysis on the feasibility and acceptability of a socially assistive robot (SAR) deployed in prospective users’ houses to monitor their cognitive capabilities through a set of digitalised neuropsychological tests and spot questions conveniently integrated within the robotic assistant’s daily tasks. We do this by describing an experimental campaign where a robotic system, integrated with a larger framework, was installed in the house of 10 users for a duration of at least 10 weeks, during which their cognitive capabilities were monitored by the robot. Concretely, the robots supervised the users during the completion of the tests and transparently monitored them by asking questions interleaved in their everyday activities. Results show a general acceptance of such technology, being able to carry out the intended tasks without being too invasive, paving the way for an impactful at-home use of SARs.

Gas source localization (GSL) with an autonomous robot is a problem with many prospective applications, from finding pipe leaks to emergency-response scenarios. In this work, we present a new method to perform GSL in realistic indoor environments, featuring obstacles and turbulent flow. Given the highly complex relationship between the source position and the measurements available to the robot (the single-point gas concentration, and the wind vector) we propose an observation model that derives from contrasting the online, real-time simulation of the gas dispersion from any candidate source localization against a gas concentration map built from sensor readings. To account for a convenient and grounded integration of both into a probabilistic estimation framework, we introduce the concept of probabilistic gas-hit maps, which provide a higher level of abstraction to model the time-dependent nature of gas dispersion. Results from both simulated and real experiments show the capabilities of our current proposal to deal with source localization in complex indoor environments.

La aparición de los modelos a gran escala permite abordar algunas de las principales limitaciones que presentan las técnicas de mapeo semántico tradicional en robótica móvil. Sin embargo, estos modelos son propensos a generar respuestas incorrectas, incoherentes o incluso inventadas, pudiendo ocasionar comportamientos erróneos del robot. Para poder desplegarse en aplicaciones reales, por tanto, es crucial desarrollar mecanismos que permitan mitigar estas errores. En este trabajo se utiliza \textit{ConceptGraphs}, un método del estado del arte basado en modelos a gran escala para construir mapas semánticos, sobre el que se plantean dos estrategias para reducir las respuestas erróneas. Primero, se propone adaptar el método para operar con modelos más recientes (por ejemplo, Gemini 1.5 y ChatGPT-4o). En segundo lugar, se incorpora una etapa de refinamiento de respuestas mediante la técnica denominada \textit{Reflexión}, que permite al modelo autoevaluar y mejorar sus propias respuestas. Finalmente, se validan las propuestas mediante experimentos en entornos reales del conjunto de datos ScanNet.

We present a novel 3D odometry method that recovers the full motion of a vehicle only from a Doppler-capable range sensor. It leverages the radial velocities measured from the scene, estimating the sensor's velocity from a single scan. The vehicle's 3D motion, defined by its linear and angular velocities, is calculated taking into consideration its kinematic model which provides a constraint between the velocity measured at the sensor frame and the vehicle frame. Experiments carried out prove the viability of our single-sensor method compared to mounting an additional IMU. Our method provides the translation of the sensor, which cannot be reliably determined from an IMU, as well as its rotation. Its short-term accuracy and fast operation ( 5ms) make it a proper candidate to supply the initialization to more complex localization algorithms or mapping pipelines. Not only does it reduce the error of the mapper, but it does so at a comparable level of accuracy as an IMU would. All without the need to mount and calibrate an extra sensor on the vehicle.

Estimating the absolute pose of a camera given a set of N points and their observations is known as the resectioning or Perspective-n-Point (PnP) problem. It is at the core of most computer vision applications and it can be stated as an instance of three-dimensional registration with point-line distances, making the error quadratic in the unknown pose. The PnP problem, though, is nonconvex due to the constraints associated with the rotation, and iterative algorithms may get trapped into any suboptimal solutions without notice. This work proposes an efficient certification algorithm for central and noncentral cameras that either confirms the optimality of a solution or is inconclusive. We exploit different sets of constraints for the rotation to assess their performance in terms of certification. Two of the formulations lack the Linear Independence Constraint Qualification (LICQ) while one of them has more constraints than variables. This hinders the usage of the ``standard"" procedure which estimates the Lagrange multipliers in closed-form. To overcome that, we formulate the certification as an eigenvalue optimization and solve it through a line-search method. Our evaluation on synthetic and real data shows that minimal formulations certify most solutions (more than 90\% on real data) whereas redundant formulations are able to certify all of them and even random problem instances. The proposed algorithm runs in microseconds for all these formulations.

This work deals with the problem of detecting and localizing methane emission sources in open spaces with a mobile robot equipped with a remote gas detector (TDLAS). To reduce the long inspection time of traditional approaches which use the ground as the natural reflector, in this work, we analyze the feasibility of a leader-follower formation, where one robot, the leader, carries the remote gas detector that scans horizontally, parallel to the ground, and a second robot, the follower, that acts as an artificial reflector. We present a visual tracking mechanism for the relative pose estimation of both mobile platforms to extend the measurement range up to 10 m. Results in a 70 m(2) experimental area demonstrate that this approach is effective for a fast location of methane gas sources.

In the Industry 4.0 era, a myriad of sensors of diverse nature (temperature, pressure, etc.) is spreading throughout the entire value chain of industries, being potentially exploitable for multiple purposes, such as Predictive Maintenance (PdM): the just-in-time maintenance of industrial assets, which results in reduced operating costs, increased operator safety, etc. Nowadays, industrial processes require to be highly configurable, in order to proactively adapt their operation to diverse factors such as user needs, product updates or supply chain uncertainties. This limits current Industry 4.0-PdM solutions, typically consisting of ad-hoc developments intended for specific scenarios, i.e. they are designed to operate under certain conditions (configurations, employed sensors, etc.), being unable to manage changes in their setup. This paper presents a general Deep Learning (DL) architecture, MachNet, which deals with such heterogeneity and is able to address PdM problems of a diverse nature. The modularity of the proposed architecture enables it to deal with an arbitrary number of sensors of different types, also allowing the integration of prior information (age of assets, material type, etc.), which clearly affects performance and is often neglected. In practice, our architecture effortlessly adapts to the assets' specifications and to different PdM problems. That is, MachNet becomes an architectural template that can be instantiated for a given scenario. We tested our proposal in two different PdM-related problems: Health State (HS) and Remaining-useful-Life (RuL) estimation, achieving in both cases comparable or superior performance to other state-of-the-art approaches, with the additional advantage of the generality that MachNet offers.