Session 6B – Robot Navigation 1
Date: Thursday, 19th November 2026
Time: 16:00 – 17:30
Room: Aula Capella
Session Chair: TBA
Presentation format: Each paper is allocated 15 minutes: approximately 10–12 minutes for the presentation and 3–5 minutes for questions.
16:00 – Cognitive Architecture for Autonomous Off-Road Driving: Planning and Decision-Making
Authors: Ramiro B. Martinez, Rodrigo Gutiérrez-Moreno, Rafael Barea, Elena López-Guillén, Kenji Sellart and Luis M. Bergasa
Affiliations: Department of Electronics, University of Alcalá (UAH), Spain.
Abstract
Off-road autonomous driving remains challenging due to terrain irregularities, degraded perception, and rapidly changing environmental uncertainty. This paper presents a hierarchical cognitive architecture for uncertainty-aware off-road navigation. The proposed framework decomposes decision-making into strategic route planning, tactical deliberation based on active inference, and low-level execution, implemented through a decentralized set of interacting agents. To support efficient real-time operation, the tactical planner relies on a compact route-relative state representation that combines local geometric features with explicit uncertainty estimates. In parallel, a dense multimodal perception branch is kept decoupled from the core tactical loop to enable diagnostic evaluation, safety auditing, and post-hoc explainability without increasing planner complexity. Experiments in an unstructured open-cast mining scenario developed in the CARLA simulator show that the architecture provides a traceable and reproducible basis for evaluating uncertainty-aware tactical behaviors in off-road autonomous navigation.
16:15 – A Quadruped-Humanoid Robotic Team for Victim Assessment in SAR Operations
Authors: Christyan Cruz Ulloa, Jaime Bravo Algaba, Fiorella Capolungo, Jaime del Cerro and Antonio Barrientos
Affiliations: Centro de Automática y Robótica (UPM-CSIC), Universidad Politécnica de Madrid, Madrid, Spain.
Abstract
Autonomous Search and Rescue (SAR) deployments demand rapid mobility and reliable preliminary victim assessment. However, single-agent perception frequently degrades under real-world conditions due to low-angle morphological constraints, environmental occlusions, and algorithmic biases. To overcome this, we propose a heterogeneous multi-robot system that tightly couples locomotion morphology with active perception. The architecture integrates an agile quadrupedal robot and a high-vantage humanoid, driven by a deep learning-based assessment pipeline. We introduce a confidence-thresholded perception strategy: the quadruped acts as a rapid ground-level scout, dynamically tasking the humanoid for unoccluded, high-angle visual verification whenever detection certainty falls below operational thresholds. The vision pipeline utilizes a CNN-seg trained on a diversity-aware hybrid dataset to explicitly mitigate demographic biases. Baseline evaluations demonstrate robust performance, achieving 97.5% accuracy for consciousness state estimation and 78.5% for injury severity classification. Since low-angle perception remains physically vulnerable to background clutter and occlusions, our multi-perspective verification resolves these spatial ambiguities. This approach effectively recovers occluded features and rejects false positives, optimizing the trade-off between search velocity and triage reliability. Datasets and code are available in Zenodo and GitHub repositories, and a demostration video is providen in the supplementary material.
16:30 – Uncertainty-Driven Disambiguation in Object-Goal Robot Navigation Using Shannon Entropy
Authors: Alicia Mora, Lucia Lishan, Noelia Fernandez and Ramon Barber
Affiliations: RoboticsLab, Department of Systems Engineering and Automation, Universidad Carlos III de Madrid, Leganés, Spain.
Abstract
Commands expressing an object goal in natural language are frequently underspecified, and environments often contain multiple matching instances. While modern object-goal navigation frameworks can locate open-vocabulary targets, they typically commit to an objective prematurely without assessing their own semantic confusion. This paper presents a framework that enables robots to evaluate their own uncertainty before commiting to a navigation goal. Operating over an episodic memory, the system couples text-embedding similarity with multimodal visual verification scores to populate a spatial belief distribution. By monitoring the Shannon Entropy of this distribution, the agent decides whether to execute navigation or to trigger dialogue interaction. The policy selects descriptive queries to break candidate ties or landmark-based questions to re-anchor unindexed targets. Real-world experiments demonstrate that the proposed method resolves ambiguities through adaptive dialogue, improving task success.
16:45 – Reactive Local Obstacle Avoidance for Autonomous Mobile Robots on Topological Maps
Authors: Marina Brilhante¹, Paulo M. Rebelo¹, Héber Sobreira¹, Fillipe L. Ribeiro² and Pedro Costa¹,³
Affiliations: ¹ INESC TEC – Institute for Systems and Computer Engineering, Technology and Science, Porto, Portugal; ² JPM Industry, Vale de Cambra, Portugal; ³ Faculty of Engineering, University of Porto (FEUP), Porto, Portugal.
Abstract
Autonomous mobile robots operating in industrial environments commonly rely on graph-based navigation systems that provide efficient and predictable route execution. However, unforeseen obstacles may temporarily block planned routes, requiring the robot to either wait or perform global replanning. While reactive obstacle avoidance is well established for metric-map navigation, comparable solutions for graph-based navigation remain limited. This paper presents an event-triggered reactive obstacle-avoidance framework that operates as a local layer on top of an existing graph-based navigation system. When a persistent obstacle blocks the planned route, the proposed method temporarily leaves the graph, generates and executes a local detour from laser observations, and explicitly rejoins the original route before resuming navigation. The framework was evaluated in simulation over 30 avoidance maneuvers in industry-inspired scenarios. In all experiments, the robot successfully bypassed the obstacles and resumed graph-based navigation without requiring global replanning. The generated detours exhibited predictable geometric and temporal overheads with low variability across repeated trials, demonstrating the feasibility of integrating reactive local obstacle avoidance into graph-based navigation.
17:00 – A Finite-State Decision-Making Framework for Autonomous Navigation in Urban Scenarios
Authors: Carolina de Diego-García, Rodrigo Gutiérrez-Moreno, Fabio Sánchez-García, Rafael Barea, Nazario Pizzicoli and Luis M. Bergasa
Affiliations: Department of Electronics, University of Alcalá (UAH), Spain.
Abstract
This paper presents the design, implementation, and evaluation of a decision-making module for autonomous navigation in urban environments, based on two coordinated finite-state machines (FSMs) and integrated into a pre-existing modular autonomous driving architecture. The proposed decision-making module acts as a high-level decision layer operating between the perception and control modules. It combines a Longitudinal Decision FSM and a Lateral Maneuver FSM under an explicit Safety Priority and Conflict Resolution Logic. Its objective is to select safe, interpretable, and easily adjustable driving behaviors. The system is implemented in Python and validated in the CARLA simulator using representative urban scenarios. The obtained results show the ability of the proposed approach to generate traceable, coherent, and safety-oriented decisions, consolidating it as an integrable solution within the architecture.
17:15 – Reward Shaping and Terrain-Relative Termination for PPO Humanoid Locomotion Beyond Flat Ground
Authors: João Viterbo Vieira¹, Francisco Pinto¹, José Pedro Carvalho² and A. Pedro Aguiar²
Affiliations: ¹ Faculty of Engineering, University of Porto, Porto, Portugal; ² SYSTEC-ARISE, Faculty of Engineering, University of Porto, Porto, Portugal.
Abstract
Humanoid locomotion is a challenging reinforcement learning problem due to high-dimensional actuation, intermittent contacts, and strong sensitivity to task and environment design. While policy-gradient methods such as Proximal Policy Optimization (PPO) reliably learn flat-ground walking, extending the same recipe to uneven terrain and long-horizon navigation frequently fails in practice, and the causes are often implementation-level rather than algorithmic. We present an empirical case study of two such design choices (reward alignment and episode-termination logic) using the MuJoCo Humanoid across four tasks of increasing difficulty: flat locomotion, stair climbing, waypoint-based circuit navigation, and circuits with stair segments. Task-aligned shaping (height progress and step milestones) yields a stair-climbing policy that reaches the top step in 96% of evaluation episodes with the smallest training budget of all four tasks. A controlled evaluation-time ablation demonstrates that fixed world-frame health bounds spuriously terminate healthy agents on elevated terrain. With terrain-relative termination, the combined navigation-and-stairs policy achieves 87% circuit completion. Environments, training configurations, and evaluation scripts are publicly released.