navigation

This work addresses the poor robustness of Simultaneous Localization and Mapping (SLAM) in dynamic, sparse environments and its heavy reliance on high-precision sensors. To this end, it systematically investigates deep integration mechanisms between reinforcement learning (RL) and SLAM. The paper introduces the first taxonomy of RL-SLAM fusion paradigms, establishes a unified evaluation framework and a technical evolution roadmap, and innovatively couples deep Q-networks, policy gradient methods, multi-agent RL, and neural SLAM architectures with graph optimization and Bayesian filtering—enabling end-to-end decision optimization and uncertainty-aware joint training with online adaptation. A comprehensive survey of over 100 studies is conducted. Experimental results demonstrate that the proposed approach improves mapping accuracy by 23% and reduces localization failure rate by 41% in dynamic sparse scenarios, while significantly enhancing generalization capability.

This paper addresses safe autonomous exploration for circular-profile robots with pose-response latency (e.g., differential-drive self-balancing robots) in unknown cluttered environments. Methodologically, it integrates IMU, 3D LiDAR, and RGB-D data within an RTAB-Map SLAM framework to achieve robust mapping and loop closure detection. A safety-aware skeleton of obstacles is constructed, where skeleton opening directions determine exploration priority, and the robot is constrained to traverse paths with high safety margins. The key contribution lies in tightly coupling geometric safety constraints—ensuring a minimum clearance from obstacles—with topological exploration guidance—driven by skeleton openings—thereby balancing collision robustness and exploration efficiency. Experimental validation in ROS-based indoor scenarios demonstrates stable, collision-free navigation, significantly improving safety-boundary maintenance and coverage rate of unknown regions.

To address the limited robustness and sensor-specific modeling dependency of LiDAR-inertial odometry under diverse sensor configurations and operational scenarios, this paper proposes a generic fusion framework that requires no prior sensor modeling. Methodologically, it employs a simplified IMU motion model for inertial integration—eliminating both feature extraction and preintegration—and introduces a direct scan-to-map LiDAR registration with a novel regularization mechanism to improve convergence stability. The key contributions are: (1) a unified configuration enabling cross-platform deployment (e.g., urban driving, natural environments) and cross-sensor compatibility (various LiDAR/IMU models); (2) experimental validation on multiple real-world robotic platforms demonstrating high accuracy, strong robustness, and real-time performance; and (3) open-sourced implementation.

This study addresses the performance evaluation of simultaneous localization and mapping (SLAM) in structured indoor environments. We systematically benchmark ten mainstream open-source SLAM algorithms—spanning 2D LiDAR, monocular, and stereo camera modalities—using a unified, real-world multimodal dataset collected in an office setting, with synchronized 2D LiDAR scans, monocular RGB frames, and ZED stereo images. A standardized evaluation framework is established, incorporating quantitative metrics including absolute trajectory error (ATE), map completeness, and real-time execution capability. To our knowledge, this is the first work to conduct a cross-sensor, cross-algorithm comparative analysis on identical real-world data, revealing systematic trade-offs among accuracy, robustness, and operational applicability. Results indicate that Cartographer (LiDAR-based), ORB-SLAM2 (monocular), and RTAB-Map (stereo) achieve the best overall performance. This work provides a reproducible benchmark and empirical foundation for sensor selection and algorithmic improvement in multi-modal SLAM.

To address the low localization accuracy and poor robustness of robot autonomous navigation in known-map environments, this paper proposes an enhanced Adaptive Monte Carlo Localization (AMCL) algorithm within the ROS framework, implemented in a multi-obstacle Gazebo simulation environment. The method fuses LiDAR measurements with motion models for pose estimation. Its core contribution is a dynamic particle count adjustment strategy that adaptively scales the number of particles based on real-time localization confidence—thereby improving computational efficiency without compromising accuracy. Experimental results demonstrate an average positional error of less than 0.15 m and an orientation error under 3° in complex scenarios. Furthermore, the approach successfully enables dual-robot navigation to multiple target waypoints. This enhancement significantly improves AMCL’s robustness and practicality under dynamic perceptual uncertainty.

To address challenges in long-duration autonomous exploration of micro air vehicles (MAVs) in unknown environments—including degradation of global consistency, insufficient safety guarantees, and poor adaptability to heterogeneous sensors—this paper proposes a subgraph-based efficient large-scale exploration framework. The method integrates visual-inertial SLAM, loop closure detection, pose-graph optimization, and sampling-based next-best-view planning, while introducing three key contributions: (1) a local subgraph model that fuses subgraph-level frontier points to generate globally consistent exploration goals; (2) an abstract multi-sensor frontend interface enabling seamless switching between LiDAR and depth cameras; and (3) tight coupling of perception, state estimation, and planning. Simulation results demonstrate a 23% improvement in exploration efficiency and an 18% gain in reconstruction accuracy over state-of-the-art methods. Real-world experiments across complex indoor and outdoor environments validate full autonomy and high trajectory consistency on two distinct MAV platforms.

This work addresses the challenge of exploration and mapping decisions in active SLAM under partial observability by formulating it as a stochastic control problem with incomplete information. The authors propose a non-standard partially observable Markov decision process (POMDP) framework that jointly integrates motion, perception, and map representation. A key innovation is the introduction of an exploration cost function that explicitly captures the geometric structure of the state space to quantify the value of information-gathering actions. Building upon this formulation, they develop a general stochastic control model and derive an approximately optimal policy with theoretical guarantees by combining stochastic control theory and reinforcement learning algorithms. Numerical experiments in representative environments demonstrate the effectiveness of the approach, successfully learning high-performance exploration strategies.

This work addresses the high communication overhead and low data efficiency in multi-robot collaborative SLAM, which often stems from reliance on low-level feature matching. The authors propose a distributed SLAM framework based on scene graph matching that leverages RGB-LiDAR fused point clouds for semantic segmentation, extracting discrete objects and their boundaries to construct scene graphs. Notably, inter-robot loop closures are achieved solely by exchanging object labels and centroids, eliminating dependence on raw feature descriptors. Integrated with a multi-stage communication scheme and distributed optimization, the method significantly reduces communication load while preserving localization and mapping accuracy, as demonstrated in both simulated and real-world experiments with legged robots across indoor and outdoor environments.

To address the need for real-time, robust LiDAR-inertial odometry (LIO) in complex environments, this paper proposes a tightly coupled LIO framework based on continuous-time B-spline trajectories. To ensure trajectory continuity and low latency, we design a non-uniform temporal node scanning-window mechanism. For efficiency and accuracy, we accelerate Gaussian mixture model (GMM) registration and covariance computation via Kronecker product decomposition, integrate unscented transform-based skew correction, and apply intra-scan piecewise motion compensation. Furthermore, IMU preintegration pseudo-measurements and relative pose soft constraints are fused into a multi-resolution surfel map joint optimization. Evaluated across handheld, ground, and aerial robot datasets, our method achieves state-of-the-art performance—improving processing speed by 3.3× while significantly enhancing robustness and real-time localization accuracy.

To address navigation challenges in partially unknown, unstructured indoor environments—such as disaster rescue scenarios—this paper proposes a real-time path planning method for detecting and actively avoiding movable obstacles. Within the ROS2 Nav2 framework, we fuse LiDAR and odometry data to construct an adaptive dynamic costmap. We introduce a dedicated cost layer for movable obstacles and a novel velocity-ratio–based slow-pose progression detection mechanism, enabling online identification of pushable obstacles and dynamic local costmap updates. The method ensures real-time performance under CPU-constrained conditions. Gazebo simulations demonstrate a significant improvement in goal attainment rate, a substantial reduction in deadlock occurrences, and traversal times comparable to baseline approaches. Moreover, the solution exhibits cross-platform deployability. Overall, it enhances the robustness of search-and-rescue robots in complex, unknown environments.

To address the insufficient localization and mapping accuracy of LiDAR SLAM in dynamic, complex environments, this paper proposes the Inferred Attention Fusion (INAF) module—a novel architecture that tightly integrates a learnable, environment-aware attention mechanism with geometric odometry for real-time co-optimization between AI models and traditional SLAM. INAF dynamically modulates multi-source feature weights based on sensor feedback, significantly improving robustness against dynamic objects, sparse structures, and illumination variations. Extensive experiments on the KITTI dataset demonstrate that INAF reduces average pose estimation error by 23.6% and improves map completeness by 18.4% compared to state-of-the-art LiDAR SLAM systems (e.g., LOAM, LIO-SAM), with particularly pronounced gains under high-speed motion and heavy occlusion. This work establishes a new paradigm for adaptive SLAM that synergistically fuses deep learning with geometric priors.