Towards Multi-Floor Autonomous Exploration with Semantic Mapping Integration

Towards Multi-Floor Autonomous Exploration with Semantic Mapping Integration Zoe Betta DIBRIS, Rice Lab University of Genova Genova, Italy [email protected] Carmine Tommaso Recchiuto DIBRIS, Rice Lab University of Genova Genova, Italy [email protected] Antonio Sgorbissa DIBRIS, Rice Lab University of Genova Genova, Italy [email protected] Abstract—This work proposes the integration of a multi-floor autonomous exploration system, based on a novel approach to frontier evaluation, with a semantic mapping system able to detect and localize relevant features using a depth camera. The system has been tested to evaluate and discuss the effect of exploration on feature detection. Index Terms—Autonomous exploration, semantic mapping, quadruped robot I. INTRODUCTION In recent years, applications in which robots increase safety in the workplace have been on the rise. For example, the Spot robot has been used in the Paris metropolitan area to inspect tunnels in place of humans [1]. Inspection tasks require the robot to cover large areas while collecting information about its surroundings. Quadruped robots are frequently used for inspection purposes due to their mobility and their ability to carry payloads. One possible application of quadruped robots to improve worker safety during inspection tasks is in port environments [2]. Workers interviewed highlighted the need for a robot to inspect and monitor harbor areas where toxic substances or radiation may be present. In this scenario, the robot is tasked with inspecting an unknown area, recording the presence of harmful gases or radiation. It is not possible to program the robot to follow the same path for each inspection due to the dynamic nature of the harbor, where containers are constantly being moved. For this application, a key requirement is the precise localization of all hazardous sources. While there are no strict time constraints, the robot must identify and localize all sources by thoroughly inspecting the area. Robots have also been employed in Search and Rescue (SAR) operations, where they can collect information while keeping first responders out of danger [3]. The trend in SAR robotics has been to use custom-built robots for specific operations. Quadruped robots have also been studied in highly hazardous terrains designed to simulate SAR conditions [4]. In our previous work, we explored a possible application of quadruped robots in SAR scenarios by interviewing first responders [5]. They emphasized the need to collect information within the time it takes them to reach the affected area. This introduces a strong constraint on the time available to complete the operation. In this case, it may be more valuable Fig. 1. System architecture and interface with the hardware to gather general information quickly rather than performing a detailed exploration of the entire space, as would be required in harbor inspections. The contribution of this work is the integration of an autonomous exploration system, capable of navigating unknown environments, with a semantic mapping algorithm that can detect relevant features such as People, Doors, Cracks, and Stairs in the environment. This system has been extensively tested to evaluate the impact of exploration strategies on feature detection. In particular, we analyze how, under time constraints, the number of detected features is influenced by the exploration strategy—whether thorough exploration or a more preliminary approach. II. METHODOLOGY The developed system is composed of two main parts: the exploration algorithm and the feature detection module. Figure 1 shows the system framework: modules related to exploration are shown in yellow, while modules for feature detection are shown in blue. The exploration component is described in our previous work [6]. It is a frontier-based algorithm with a novel frontier evaluation method that considers the estimated unexplored area on each floor. Frontiers are assessed based on the expected information gain from reaching them. The exploration behavior can be regulated by two parameters: the floor weight Fw, which determines whether the robot prefers to fully explore a floor before moving to the next one (lower values favor full2025 I-RIM Conference October 17-19, Rome, Italy ISBN: 9788894580570 10.5281/zenodo.17629816 185 TABLE I EXPERIMENTAL RESULTS FwOptimization People Floor visited out Area explored 0.3 ON 9 15 709.39 m2 0.3 OFF 7 13 535.55 m2 0 ON 9 12 591.93 m2 1 ON 7 16 635.76 m2 floor exploration, while higher values favor switching floors earlier), and the optimization parameter, which can be toggled on or off to encourage the robot to select straighter paths aligned with its previous direction of motion. The generation of semantic maps is described in previous work [7]. Features are detected using a YOLO-based model and are projected onto a map, and updated using the same principle as a Probabilistic Occupancy Grid Map [8]. This enables the merging of detections that are spatially close and the aggregation of repeated detections. III. EXPERIMENTS To evaluate the performance of the system, a series of structured experiments was conducted to assess both the exploration and recognition components. The aim was to determine whether there is a difference between thoroughly exploring each floor before moving to the next one or quickly scanning each floor before proceeding. The objective was not only to assess exploration performance but also to determine whether the number of detected objects changes depending on the exploration strategy. In scenarios with strict time constraints, one behavior may be more effective, whereas in situations without such constraints, the priority is to collect all information as precisely as possible. The robot was operated fully autonomously for 10 minutes in each trial, conducted on the first floor of the B building of the DIME Department at the University of Genoa. We tested the system with four sets of parameters (listed as rows in Table I), each repeated across five configurations in which the robot started from a different floor and with people positioned in different locations. Each configuration included four floors with one person per floor. For each parameter set, the total number of people to detect was 20 (four per trial), and the total number of floors to visit was 20 (four per trial). To ensure consistency, whenever the robot changed floors, a fixed duration of one minute was assigned for the stair-climbing. Table I reports the experimental results. The third column shows the number of people detected under each parameter set. The differences across parameter sets are not large, but a more detailed discussion emerges when considering the number of floors visited and the total area explored. The first row corresponds to a balanced configuration with optimization enabled. The second row shows the same balanced configuration with optimization disabled. The third row represents the exhaustive configuration, in which the robot prefers to thoroughly explore each floor. The fourth row represents the greedy configuration, in which the robot strongly favors changing floors as soon as possible. Notably, the second row (optimization disabled) exhibits the worst performance, both in terms of explored area and People detected. Comparing the balanced and exhaustive configurations, we observe that both detected the same number of people. However, the balanced configuration explored a larger area. This is because the balanced robot quickly explores part of each floor before moving on and, upon reaching a new floor, immediately discovers large unexplored areas. The exhaustive robot remains on a single floor until it is fully explored, discovering progressively smaller new areas as time passes. When comparing the balanced and greedy configurations, we see that the greedy robot visited more floors, yet covered a smaller overall area. This is because changing floors provides an immediate gain in newly explored area but also consumes time for the transition. Moreover, the greedy configuration detected fewer people, since the robot mainly focused on the areas near the stairs before switching floors. In conclusion, the choice of parameters has a significant effect on the robot’s behavior and performance. Selecting the most suitable parameters depends on the application’s requirements, particularly regarding time constraints and the level of detail needed. In some cases, approximate exploration may be sufficient to provide a general overview, while in others, precise and thorough exploration is necessary to ensure that no features are missed. IV. CONCLUSION This work proposes a system that integrates autonomous exploration in multi-floor buildings with the detection and localization of relevant environmental features. The system was tested to evaluate performance differences across parameter sets, and the influence of these parameters on the robot’s behavior is also discussed. REFERENCES [1] (2023) Ratp case study, spot boston dynamics. [Online]. Available: https://www.bostondynamics.com/resources/case-study/ratp [2] Z. Betta, D. Corongiu, C. T. Recchiuto, and A. Sgorbissa, “Enhancing worker safety in harbors using quadruped robots,” in WIP 22nd Int. Conf. Ubiquitous Robots (UR 2025), College Station, TX, USA, Jul 2025, pp. 1–10. [3] R. R. Murphy, Disaster Robotics. The MIT Press, 2014. [4] I. D. Miller, F. Cladera, A. Cowley, S. S. Shivakumar, E. S. Lee, L. JarinLipschitz, A. Bhat, N. Rodrigues, A. Zhou, A. Cohen, A. Kulkarni, J. Laney, C. J. Taylor, and V. Kumar, “Mine tunnel exploration using multiple quadrupedal robots,” IEEE Robot. Autom. Lett., vol. 5, no. 2, pp. 2840–2847, 2020. [5] Z. Betta, A. Gaudino, A. Benini, C. T. Recchiuto, and A. Sgorbissa, “Perceptions and opinions of rescuers about a quadruped robot in an earthquake scenario,” in 2024 33rd IEEE International Conference on Robot and Human Interactive Communication (ROMAN), 2024, pp. 1092– 1099. [6] Z. Betta, S. Paneri, A. Gaudino, A. Benini, C. T. Recchiuto, and A. Sgorbissa, “Multi-floor danger and responsiveness assessment with autonomous legged robots in catastrophic scenarios,” in IEEE RoMan 2023, 2023. [7] Z. Betta, C. T. Recchiuto, and A. Sgorbissa, “People, cracks, stairs, and doors: vision-based semantic mapping with a quadruped robot supporting first responders in search & rescue,” in 2024 33rd IEEE International Conference on Robot and Human Interactive Communication (ROMAN), 2024, pp. 1878–1885. [8] A. Elfes, “Using occupancy grids for mobile robot perception and navigation,” Computer, vol. 22, no. 6, pp. 46–57, 1989. 186