A Hybrid Learning - Augmented Monte Carlo Framework for Mobile Robot Localization in a Symmetrical Environment

Shiron Anto Melvin

doi:10.60867/00000180

Autonomous mobile robots have significant potential to improve efficiency, productivity, and sustainability in modern greenhouse operations. However, reliable navigation in greenhouse environments remains particularly challenging because the highly repetitive arrangement of crop rows is combined with occlusion and changing plant geometry, resulting in LiDAR observations that can appear locally similar across spatially separated locations. Unlike structured environments such as warehouses, where fixed rack geometries and relatively stable layouts can provide persistent geometric features for scan matching and map-based localization, greenhouse environments contain living and evolving vegetation whose appearance and geometry vary with crop growth, cultivation activities, and seasonal conditions. These characteristics increase perceptual ambiguity and make reliable global localization from geometric observations alone difficult, particularly when the robot has an uncertain or completely unknown initial pose. Among probabilistic localization approaches, Monte Carlo Localization (MCL) provides a suitable framework for this problem because it explicitly represents multiple competing pose hypotheses and can therefore accommodate global localization and multimodal uncertainty, which are difficult to represent with locally linear estimators such as the Extended or Unscented Kalman Filter. However, conventional MCL relies primarily on geometric consistency between LiDAR observations and the map and can therefore struggle when perceptually similar observations produce comparable likelihoods for multiple locations. This thesis addresses these challenges by developing a learning augmented probabilistic localization framework that combines deep learning with MCL to achieve robust, real-time global localization using only a low-cost 2D light detection and ranging (LiDAR) sensor. The motivation for incorporating learning is that, although greenhouse LiDAR observations may appear geometrically ambiguous at a global level, they can contain subtle spatially discriminative patterns arising from differences in crop row structure, vegetation distribution, and local environmental geometry. A learned model can capture such patterns directly from LiDAR observations and provide an additional pose-related cue to complement the map-based likelihood used by MCL. The proposed approach, therefore, retains the multimodal hypothesis representation and probabilistic state estimation capability of MCL while augmenting its geometric measurement model with learned information. The resulting framework is designed to improve the localization system's ability to distinguish perceptually similar locations and recover the correct global pose under substantial initial pose uncertainty, while maintaining the computational efficiency required for real-time robot operation. To establish the research foundation, a comprehensive review of mobile robot localization techniques is conducted. Classical probabilistic localization methods, simultaneous localization and mapping (SLAM) approaches, optimization-based techniques, and recent learning-based localization methods are systematically examined. Particular attention is given to the global localization problem in structured and symmetrical environments, where perceptual aliasing and multi-hypothesis ambiguity significantly degrade localization performance. The review identifies a key research gap in the lack of effective integration between learning-based pose estimation and probabilistic localization frameworks for low-cost sensing systems operating in greenhouse environments. This gap motivates the development of a hybrid localization approach that leverages both data-driven perception and probabilistic reasoning. As an initial feasibility study, a Deep Monte Carlo Localization (DMCL) framework is developed and evaluated in a simulated greenhouse environment. A LiDAR preprocessing pipeline is proposed to convert raw 2D laser scans into structured spatial projection images suitable for supervised learning. A convolutional neural network is then designed to predict robot pose directly from the generated LiDAR images. The predicted pose information is incorporated into the Monte Carlo Localization framework via a hybrid observation model that combines a learning-based pose likelihood with a conventional LiDAR sensor likelihood. Simulation results demonstrate that integrating learning-based information significantly improves ambiguity resolution and global localization performance in symmetrical environments. The proposed framework achieves a global localization success rate of 94% compared to 4% for standard MCL, reduces trajectory pose error by up to 97%, and converges in less than one second. The hybrid observation model does not represent conventional multi-sensor fusion, as both the conventional sensor likelihood and the learning-based pose likelihood are derived from the same 2D LiDAR measurements. The conventional LiDAR likelihood evaluates the geometric agreement between the observed scan and the environment map, whereas the learning-based pose likelihood provides an additional data-driven estimate of the robot pose from the LiDAR observation. These two components, therefore, provide complementary interpretations of the same sensor data, with the learned component capturing spatially discriminative patterns that may be difficult to distinguish using geometric map matching alone. Integrating these complementary likelihoods within the MCL framework enables the probabilistic localization process to benefit from learned spatial information while retaining the ability to represent and evaluate multiple pose hypotheses. This motivates the development of a hybrid localization approach that combines data-driven pose estimation with probabilistic reasoning to address perceptual ambiguity in greenhouse environments. However, the study also reveals important limitations, including a limited feature representation arising from the single-channel LiDAR image, computational inefficiency due to the CNN architecture, and the limited effectiveness of the initial weighting strategy. These limitations motivate further improvements toward practical real-time deployment. To address the representational and computational limitations identified in the initial framework, a novel dual-channel LiDAR representation and a lightweight convolutional neural network architecture, termed DualLiDARNet, are proposed. The dual-channel representation combines spatial projection information with normalized raw LiDAR range measurements within a 320 × 320 image, enabling the network to learn geometric occupancy patterns and range-based structural features simultaneously. The 320 × 320 resolution was selected as a practical operating point that provides sufficient spatial discretization of the LiDAR measurements to preserve location-relevant geometric features while avoiding unnecessary increases in computational and memory requirements associated with higher resolution representations. A lightweight architecture is specifically designed to balance localization accuracy and computational efficiency for real-time robotic applications. Extensive evaluation is conducted using multiple backbone architectures, including ResNet and MobileNet, to analyze the trade-off between accuracy and inference cost. Experimental results demonstrate that DualLiDARNet achieves a mean position error of approximately 0.07 m and an orientation error of 2.5°, while maintaining an inference time of approximately 85 ms per frame. At this inference rate, the network can provide pose estimates at approximately 11.8 Hz, which is compatible with the greenhouse robot's low-speed operation and provides sufficiently frequent pose updates for continuous localization during navigation. The achieved position accuracy is also small relative to the spatial constraints of greenhouse navigation, where maintaining reliable separation from crop rows is important for safe robot operation. Therefore, the achieved accuracy and computational performance provide an appropriate operating point for the intended greenhouse application rather than representing an unnecessarily high-accuracy or computationally overdesigned solution. The proposed dual-channel representation and lightweight architecture significantly improve pose estimation performance and provide an efficient learning model suitable for integration into a real-world localization framework. Building on the improved pose estimation capability of DualLiDARNet, a refined learning-augmented Monte Carlo Localization framework is developed for real-world deployment. A hybrid observation model is formulated that combines a conventional map-based scan-matching likelihood with learning-based pose-prediction information. To effectively resolve localization ambiguity in symmetric environments, a penalty-based hybrid weighting mechanism is introduced that adjusts particle weights based on their spatial and angular deviations from the predicted pose. This approach suppresses semantically inconsistent hypotheses while preserving probabilistic consistency and particle diversity. The resulting DMCL framework is integrated into a complete greenhouse robot navigation system and deployed on a custom-built robotic platform operating in real greenhouse environments. Experimental validation demonstrates that the proposed framework achieves localization root-mean-square errors below 0.12 m and angular errors below 0.05° while operating solely with a low-cost 2D LiDAR sensor and without requiring cameras, artificial landmarks, or additional infrastructure. The proposed approach successfully resolves multi-hypothesis ambiguity, improves convergence speed, and enhances localization robustness in highly symmetrical greenhouse environments. By bridging the gap between data-driven learning and probabilistic localization, this research provides a scalable, infrastructure-free, and practically deployable solution for autonomous greenhouse robotics, advancing reliable navigation systems for future agricultural automation.

A Hybrid Learning - Augmented Monte Carlo Framework for Mobile Robot Localization in a Symmetrical Environment

Files and links (1)

Abstract

Details

Metrics