Egocentric Blind Mapless Navigation for Cognitive Robotics

Egocentric Blind Mapless Navigation for Cognitive Robotics 1st Giovanni De Gasperis DISIM University of L’Aquila L’Aquila, Italy [email protected] 2nd Sante Dino Facchini DISIM University of L’Aquila L’Aquila, Italy [email protected] 3rd Luigi Iaconi DISIM University of L’Aquila L’Aquila, Italy [email protected]vaq.it Abstract—This article presents an egocentric mapless navigation approach for cognitive robotics that avoids pre-built maps and explicit localization. Using LiDAR embeddings and vector search, the system learns navigation commands through humanguided training and adapts in real time via cosine similarity comparing real-time measurements with the vector memory. A FlowProtocol-based architecture with Moliris and SkRobot enables distributed control, supporting robust, efficient, and adaptive autonomous navigation in dynamic environments. Index Terms—Map less navigation, Cognitive Robotics, Realtime DDS I. MAPLESS APPROACH FOR ROBOT NAVIGATION Navigation functionalities are essential in robotics, especially when energy consumption is a primary constraint. Moving without relying on an externally-built geometric map or explicit global localisation, while leveraging onboard sensing, learned policies, and internal representations to perceive, decide and act could be some essential features [1], similar to how a human navigates in an unknown environment. Such a map-less configuration, in fact, integrates with Cognitive AI paradigms like embodied perception, adaptive decisionmaking, and compact internal representations, allowing elimination of energy-consuming hardware and retrainings [2]. One main research direction in the literature is end-to-end learning and deep Reinforcement Learning (RL). Agents learn policies directly from raw sensors to reach targets and avoid obstacles, often augmented with auxiliary tasks and memory modules to improve data efficiency and generalisation. Examples include [3] and target-driven visual navigation approaches [4]. The other main direction is hybrid or Neural Simultaneous Localisation and Mapping (SLAM) approaches, where learned components build internal spatial representations (external memory, attention mechanisms) that play the role of a “soft” map while still avoiding explicit metric mapping like Neural SLAM and related memory-augmented agents [5]. This work was supported by ”Bando a cascata Spoke 6 - MOST-UniMore” funded by the European Union – NextGenerationEU, within the Italian National Recovery and Resilience Plan (PNRR), under the grant IMPACT (CUP: E93C22001070001). II. METHODS AND FRAMEWORKS (i)Agilex Scout Mini1is a compact mobile robotic platform featuring four independent drive wheels with differential steering that enables zero-radius turning. Equipped with a highperformance LiDAR sensor and Jetson Nano board, it runs the Moliris robot management system. Moliris offers ROS and CAN bus compatibility, enabling development of a web application that creates environmental maps and allows remote control through a virtual gamepad interface. (ii)Realtime Points Cloud Generation: The Lidar reader class extracts raw LiDAR data via Python websocket library from Moliris middleware, generating point clouds with dual functionality. During training phases, point clouds containing up to 360 coordinate points are stored as embeddings in a chromadb vector database alongside metadata containing executed commands for specific situations. During autonomous navigation, newly acquired point clouds are compared against training data using cosine similarity to identify optimal command execution patterns for obstacle avoidance and smooth environmental exploration. (iii)Flowprotocol and SkRobot is a Distributed Data System (DDS) initially exposed in [6], designed to interconnect heterogeneous autonomous entities (human and nonhuman). FlowNetwork supports centralised (hub–satellite) and peer-to-peer (P2P) configurations through the FlowProtocol, enabling synchronous and asynchronous communication with efficient responsiveness, parallel execution, real-time operation, and event/data distribution. In centralised deployments, SkRobot2instances function as hubs, orchestrating lightweight Python sketches executed on satellites III. ARCHITECTURE: FUNCTION-CALLING AGENT The system’s architecture leverages a Micro-agent paradigm where nodes call session instances and scripts to act as micro agents using Moliris as a middle layer. We implemented an API using the two scripts “moliris body.py” and “lidar reader.py” to achieve data handling and real-time communication, focusing on LiDAR data acquisition and management, virtual control, and command logging. Websockets connect to a remote endpoint that streams LiDAR data in real 1https://global.agilex.ai/products/scout-mini, accessed Oct 2025 2https://gitlab.com/Tetsuo-tek/SkRobot, accessed Oct 2025 2025 I-RIM Conference October 17-19, Rome, Italy ISBN: 9788894580570 10.5281/zenodo.17629856 217 time. In Figure 1, we represent a scheme of the FlowProtocol nodes and the mutual interaction among entities composing the system. We implement a Macro-agent that supervises the system behaviour, involving a set of micro-agents as executors. Fig. 1. Micro-agent nodes (orange) and class instances (sky blue) interact to form a Macro-agent that navigates after a supervised training phase. Scripts (violet) implement human controller GUI and WASD interactions. IV. EMBEDDINGS AND VECTOR-COMMAND COUPLING This approach allows for flexible training of an autonomous navigation system, without relying on prior mapping of the environment. Navigation is based on the recognition of perceived scenarios in real time, associating them with previous situations, rather than on pre-existing knowledge of the environment map. LiDAR embeddings associated with specific commands enable inference based on sensory similarity without explicit localization. This method is valuable in dynamic or non-mappable contexts, expanding the application possibilities in autonomous robotics. It has two main characteristics: (i) Training and dataset creation. Training is performed under the control of a human operator who must carefully explore the surrounding environment using the WASD keys on their keyboard. Subsequently, the point cloud is saved as a CSV file with the selected navigation command (Forward, Backward, Left, and Right). The point cloud is then transformed into an embedding usingChromaDB, obtaining a vector composed of up to 360 LiDAR points structured as a vector associated, as metadata, with the command used in that situation. The model is built as a union of integrated data from multiple environments, making the system more robust and adaptable. (ii) Navigation and creation of historical datasets. During navigation, the algorithm acquires real-time environmental data from the LiDAR sensor, transforming it into spatial representations. This data is then compared with the information stored in the vector archive, using a metric based on cosine similarity. Li et al [7] have confirmed our experience that a comparison is valid if the cosine similarity value is greater than or equal to 0.6, the is considered sufficient to indicate a meaningful match between high-dimensional vectors in many practical applications. When the algorithm finds an acceptable match it proceeds with the next operational decision. Furthermore, to prevent deadlock, we extended the navigational algorithm by calculating the exponential moving average (EMA) of the sensory vector. This approach, when compared with the raw current data using cosine similarity, enables the mobile rover to recognise deadlock situations. Therefore, the system issues a random command between right, left, and back, thereby allowing the robot to exit the deadlock position. After this deviation, the similarity-based comparison cycle restarts to validate the new direction. This mechanism prevents the robot from getting stuck and maintains adaptive and consistent control with the environment. The code used in the present work is visible at our Github repository3while the dataset and other experimental items are available on Zenodo4 V. CONCLUSIONS AND FUTURE WORKS The real innovation lies in the fact that the robot’s position is not known during training, which makes it impossible to create an accurate LiDAR map. Each training phase represents a form of reinforcement, meaning that we may simultaneously have data collected from different environments in the final dataset. This egocentric approach allows obstacle avoidance without relying on absolute position on a map and without odometry data. It puts the robot at the centre of the decision process, relying only on its environmental perceptions. As future development a generalization within the robotic decision-making process could be improved by substituting the fixed threshold mechanism with a lightweight, learned confidence estimator. In particular, a compact classification head, operating on the retrieved vector representation, could dynamically assess the reliability of the selected command with respect to the current perceptual and task context. This mechanism would enable more adaptive and context-aware command selection. ACKNOWLEDGMENT AND DECLARATIONS Authors occasionally used generative AI tools to improve the readability of some parts of the text, they reviewed the content and took full responsibility for the paper’s content. REFERENCES [1] S. Nahavandi, R. Alizadehsani, D. Nahavandi, S. Mohamed, N. Mohajer, M. Rokonuzzaman, and I. Hossain, ”A Comprehensive Review on Autonomous Navigation,” ACM Comput. Surv. 57, 9, Article 234, September 2025, 67 pages. https://doi.org/10.1145/3727642 [2] I. Polykretis and A. Danielescu, ”Mapless mobile robot navigation at the edge using self-supervised cognitive map learners,” Front Robot AI, 2024 May 22;11:1372375, doi: 10.3389/frobt.2024.1372375. [3] P. Mirowski, ”Learning to Navigate in Complex Environments,” arXiv e-prints, arXiv-1611, 2016. [4] Y. Zhu et al., ”Target-driven visual navigation in indoor scenes using deep reinforcement learning,” 2017 IEEE International Conference on Robotics and Automation (ICRA), Singapore, 2017, pp. 3357-3364, doi: 10.1109/ICRA.2017.7989381. [5] J. Zhang, L. Tai, M. Liu, J. Boedecker, and W. Burgard, ”Neural slam: Learning to explore with external memory,” arXiv preprint arXiv:1706.09520, 2017. [6] G. De Gasperis, D. Di Ottavio & S. D. Facchini, ”SkRobot with TeleoR/QuLog: A Pseudo-Realtime Robotics Data Distribution Service Extended with Production Rules and Reasoning,” Proc. 21st Int. Conf. Informatics in Control, Automation and Robotics, vol. 1, pp. 408-415, 2024. DOI: 10.5220/0013014900003822. [7] Li, Y., Wang, J., Pullman, B. et al. Index-based, High-dimensional, Cosine Threshold Querying with Optimality Guarantees. Theory Comput Syst 65, 42–83 (2021). https://doi.org/10.1007/s00224-020-10009-6 3https://github.com/IMPACT-UnivAQ/mapless-navigator, Oct 2025 4https://doi.org/10.5281/zenodo.17186863, accessed Oct 2025 218