cs.ROApr 14, 2025
SaveRegion Based SLAM-Aware Exploration: Efficient and Robust Autonomous Mapping Strategy That Can Scale
Organizations: Amazon Lab126
Abstract
Autonomous exploration for mapping unknown large scale environments is a fundamental challenge in robotics, with efficiency in time, stability against map corruption and computational resources being crucial. This paper presents a novel approach to indoor exploration that addresses these key issues in existing methods. We introduce a Simultaneous Localization and Mapping (SLAM)-aware region-based exploration strategy that partitions the environment into discrete regions, allowing the robot to incrementally explore and stabilize each region before moving to the next one. This approach significantly reduces redundant exploration and improves overall efficiency. As the device finishes exploring a region and stabilizes it, we also perform SLAM keyframe marginalization, a technique which reduces problem complexity by eliminating variables, while preserving their essential information. To improves robustness and further enhance efficiency, we develop a checkpoint system that enables the robot to resume exploration from the last stable region in case of failures, eliminating the need for complete re-exploration. Our method, tested in real homes, office and simulations, outperforms state-of-the-art approaches. The improvements demonstrate substantial enhancements in various real world environments, with significant reductions in keyframe usage (85%), submap usage (50% office, 32% home), pose graph optimization time (78-80%), and exploration duration (10-15%). This region-based strategy with keyframe marginalization offers an efficient solution for autonomous robotic mapping.
Figures & tables
Fig. 1 : Region based SLAM-aware (R-ALC) Exploration . Our strategy explores the environment region by region. Planners ensure each region is fully explored and stabilized before moving to the next, preventing unnecessary back-and-forth movement during exploration, and improving the efficiency. Light green rectangles indicate regions under exploration, while dark green ones show completed regions. Once a region is completed, the SLAM system freezes its poses for future optimization, preventing relative motions within this region while allowing global transformations. This approach significantly reduces computational and memory requirements. In the case of exploration failure, the system loads the latest completed region, enabling resumption from a stable state.
Fig. 2 : Case study of region based exploration process simulated in HSL#1. (a) We show Region 1 (a light green rectangle) that is currently undergoing region discovery phase. The robot focuses on addressing all frontiers within Region 1, while ignoring those outside Region 1. (b) As Region 1 enters the refinement phase, the PGS planner generates a convex hull (a light green polygon) using the keyframes of the pose graph inside the current region. Then the robot navigates along this convex hull in both clockwise and counter-clockwise directions to ensure regional pose graph stability. (c) The ALC planner decides an ALC target (a large green cube) from the set of ALC candidates (small green cubes), all situated within Region 2. The ALC planner is activated as uncertainty reduction exceeds the threshold, and drives the robot to the ALC target to attempt a loop closure and improve pose graph local stability. (d) Upon completion of all regions (both discovery and refinement phases), the PGS planner computes a convex hull using all the keyframes on the global pose graph. The robot then traverses all vertices of this convex hull to ensure global stability.
Fig. 3 : System architecture: The Map Checkpoints block stores the map checkpoint after each successful explored region. The Region Manager manages the region information. The Mapping component handles current map management and pose graph optimization/marginalization. The Exploration component handles the whole exploration logic using current map data and region information. This component uses three planners: 1) Frontier planner determines the next frontier to explore, 2) Active Loop Closure (ALC) planner performs active loop closure, 3) Pose Graph Stabilizing (PGS) planner stabilizes the pose graph. Finally, the Robot Planner component executes the motion to the next global plan determined by the Exploration component.
Fig. 4 : Region Based Exploration Flow Chart. During each exploration planning cycle at a fixed rate, we begin with the region discovery phase, alternating between ALC and frontier planners. After the region is fully discovered, we enter the region refinement phase using the PGS planner. After a region is completed (i.e. discovered and stabilized), we create a new one, and repeat the process again. When all regions are completed, we run the PGS planner on global scale to stabilize the global map, concluding the exploration.
Fig. 5 : Pose Graph Marginalization over a Region. The left image shows the pose graph after exploring and stabilizing a region (grey box), with keyframes represented as blue and red dots and factors as black edges. A subset of keyframes (of size 1 for demonstration purposes here), represented as a red dot, is selected within the region to perform pose graph marginalization. The marginalization process removes all blue keyframes while retaining the information of the factors in new connections. The right figure displays the pose graph after marginalization, which includes only the single anchor keyframe within the region and new factors (green edges) connecting the remaining keyframe to keyframes outside of the region. This process significantly reduces the complexity of the pose graph while preserving essential spatial information.
Fig. 6 : Comparison of exploration duration (a), submap count (b), keyframe count (c), and average pose graph optimization time (d) between R-ALC, R-ALC without keyframe marginalization, and the baseline ALCv2 [ 3 ] across different simulation environments (HSL#1 and HSL#2) and real-world environments (Home and Office Env). Lower values are better for all plots. R-ALC reduces exploration duration (by in office, in home) and generates a global occupancy map with fewer submaps ( fewer in office, fewer in home). Additionally, R-ALC uses significantly fewer keyframes ( fewer) to construct the SLAM pose graph, resulting in a substantial decrease ( ) in the average pose graph optimization time.
Fig. 7 : Example of an exploration recovery scenario from a real device test. The figures visualize the global occupancy map of the environment as it is being generated during an exploration session. Black cells indicate occupied areas, blue represents free space, and yellow shows the robot’s trajectory (dilated by its radius). Green rectangles denote exploration regions. a) The device begins exploration. b) It encounters a failure while exploring the second region due to inability to find a path to a previously visited location. The highlighted area is blocked by simulated clutter, representing potential map distortions caused by false obstacle detection and/or localization drift. c) The device is returned to its dock, and exploration is restarted. The system restores the latest saved progress, corresponding to the map state at the completion of exploration of the first region. d) The device explores the remaining unvisited areas and completes the exploration task.
Fig. 8 : Comparison of the quality of occupancy maps generated by R-ALC and the baseline exploration algorithms, evaluated using the aggregate mean IoU (higher is better) and mean DTE (lower is better) across all exploration runs in two simulation environments. The IoU results show comparable alignment between the reference and test occupancy maps for all three exploration methods. The DTE metric results show that the occupancy map errors between the exploration methods differ by less than , which is smaller than the grid cell size threshold of .
Fig. 9 : Visualization of the occupancy maps generated after autonomous exploration with R-ALC overlaid on the ground truth maps for the 9(a) HSL#1 and 9(b) HSL#2 simulation environments. The occupied cells in the test and reference maps are visualized in blue and green, respectively. The generated maps are globally consistent with the reference maps.