In the current era, with the rapid advancement of robotics and artificial intelligence, embodied intelligence has ushered in a new wave of developmental opportunities. Embodied intelligence, which integrates AI into physical entities like robots, enables these agents to interact with the environment akin to humans, acquiring information, understanding problems, making decisions, and executing tasks to produce intelligent and adaptive behaviors. This significantly enhances the intelligence and motion capabilities of robots. Consequently, a diverse array of embodied AI robots is continually emerging and is gradually being applied across various fields such as construction, security patrols, emergency response, and industrial production, steadily improving quality of life. It is evident that embodied AI robots are a focal point of contemporary research. The ability of an embodied AI robot to navigate autonomously and perform tasks in real-world scenarios hinges critically on two conditions: constructing high-precision maps and achieving real-time localization. High-precision maps form the foundation for an embodied AI robot’s navigation and task planning, while real-time localization refers to the robot’s ongoing estimation of its own pose within the environment.
Simultaneous Localization and Mapping (SLAM) technology provides embodied AI robots with both high-precision maps and real-time pose estimation. SLAM enables an embodied AI robot to move autonomously in an unknown environment without a prior map, using sensors to gather information about its surroundings, estimate its position, and concurrently create a map of the environment. Based on the primary sensor used, SLAM is broadly categorized into Visual SLAM and LiDAR SLAM. Visual SLAM relies on cameras as the core sensor, using image information for localization and mapping. Despite significant progress, Visual SLAM is highly sensitive to lighting conditions, and its performance degrades substantially in low-light environments. In contrast, LiDAR SLAM primarily utilizes LiDAR sensors to obtain high-precision distance information. It is unaffected by lighting conditions and can achieve high-precision localization and mapping in complex environments such as construction sites.
Most mainstream 3D LiDAR SLAM algorithms are predicated on the assumption of a static environment. However, practical application scenes like construction sites, urban roads, and large shopping centers often contain numerous dynamic objects such as pedestrians, vehicles, and animals. The presence of these dynamic objects creates ‘ghost’ trails or artifacts in the 3D point cloud map. On one hand, these artifacts can be mistakenly identified as obstacles, interfering with subsequent path planning for the embodied AI robot, and they may also occlude static objects, reducing map accuracy. On the other hand, since 3D LiDAR SLAM estimates its pose through point cloud registration, using point clouds containing these artifacts introduces errors in pose calculation, leading to a decline in the robot’s localization accuracy. Therefore, effectively removing dynamic points to enable accurate localization and the construction of precise maps has become a prominent research direction in the field of embodied AI robots. This review aims to summarize and discuss relevant research on 3D LiDAR SLAM for embodied AI robots in dynamic environments, covering methodological approaches, processing strategies, evaluation metrics, and future trends.
Methodologies for Dynamic Point Cloud Removal
The core challenge in dynamic environment SLAM is identifying and filtering out points belonging to moving objects before they corrupt the map or the pose estimate. Based on the underlying detection principle, traditional dynamic point removal methods can be classified into three main categories.
1. Semantic Segmentation-Based Methods
These methods typically employ clustering or, more recently, deep learning techniques to identify and segment object classes like pedestrians and vehicles within the point cloud. The fundamental principle is that if common dynamic object classes can be accurately recognized, their corresponding points can be easily eliminated. Early works used geometric modeling, such as 2D bounding boxes for vehicles. With the advent of deep learning, networks like FlowNet3D pioneered end-to-end scene flow estimation from two point clouds, which can be used for motion segmentation. Subsequent networks like SalsaNext introduced uncertainty-aware semantic segmentation for robust understanding. To process points directly in 3D space preserving geometric relationships, Cylinder3D was developed using asymmetric residual blocks and dimension-decomposed context modeling. Methods like RVMOS combine motion features with environmental semantics to identify both moving and temporarily static dynamic objects. A trend is towards multi-modal fusion, as seen in MotionSeg3D, which fuses appearance features from range images with temporal motion features from residual images. For large-scale scenes, methods like MINet and polar cylindrical balanced random sampling aim to balance efficiency and effectiveness.
The general pipeline involves passing LiDAR scans through a trained neural network (e.g., RangeNet++, Cylinder3D) to obtain per-point semantic labels. Points labeled as dynamic classes (person, car, bicyclist, etc.) are then filtered out. A key metric for segmentation performance is the Intersection over Union (IoU). For a class \(c\), it is calculated as:
$$IoU_c = \frac{TP_c}{TP_c + FP_c + FN_c}$$
where \(TP_c\), \(FP_c\), and \(FN_c\) are true positives, false positives, and false negatives for class \(c\), respectively. The mean IoU (mIoU) across all classes is a standard benchmark.
| Year | Author(s) | Key Contribution | Limitation |
|---|---|---|---|
| 2009 | Petrovskaya et al. | 2D bounding box modeling for vehicles. | Poor applicability to pedestrians, cyclists. |
| 2018 | Ruchti et al. | Estimating point-wise dynamic probability via neural network. | Cannot detect untrained object types. |
| 2019 | Liu et al. | FlowNet3D for end-to-end scene flow estimation. | Does not integrate motion priors. |
| 2020 | Zhou et al. | Cylinder3D for 3D topology-aware segmentation. | High computational cost, not real-time. |
| 2022 | Kim et al. | RVMOS fusing motion and semantic features. | Difficulty detecting small dynamic objects. |
| 2022 | Sun et al. | MotionSeg3D with motion-guided attention. | Performance drops with extra training data. |
While effective for known classes, these methods rely heavily on large, annotated datasets, can only recognize pre-trained types, demand significant GPU resources, and often suffer from poor generalization, potentially missing parts of objects.
2. Ray Tracing-Based Methods
This approach discretizes the 3D space into voxels and uses ray casting from the LiDAR’s perspective to reason about occupancy over time. The core principle is that dynamic points occupy voxels only transiently; a laser ray will hit that voxel briefly when the object is present but will pass through it at other times. By maintaining a probabilistic count of hits per voxel (e.g., in an OctoMap), voxels with a hit count below a certain threshold are considered dynamically occupied, and points within them are removed.
Early work by Underwood et al. performed ray tracing in spherical coordinates for change detection. The OctoMap framework provided a foundational probabilistic octree-based mapping system. Methods like PeopleRemover used a voxel grid with ray intersection tests and introduced a ‘safety cutoff’ to reduce errors from large incidence angles. Subsequent approaches often incorporate ground segmentation as a preprocessing step to simplify the dynamic object detection problem on non-ground points. The probability \(P(n|z_{1:t})\) of a voxel being occupied after \(t\) observations \(z\) can be updated using a log-odds formulation:
$$L(n|z_{1:t}) = L(n|z_{1:t-1}) + L(n|z_t)$$
where \(L(n) = \log \left( \frac{P(n)}{1-P(n)} \right)\). A voxel is considered free if \(L(n)\) is below a threshold \(l_{free}\) and occupied if above \(l_{occ}\). Dynamic voxels are those inconsistently observed.
| Year | Author(s) | Key Contribution | Limitation |
|---|---|---|---|
| 2013 | Hornung et al. | OctoMap: Probabilistic 3D mapping using octrees. | Trade-off between voxel size and computational load. |
| 2018 | Schauer et al. | PeopleRemover with safety cutoff mechanism. | High memory and computational consumption. |
| 2020 | Pagad et al. | Occupancy map construction using octrees. | Not suitable for very large-scale outdoor maps. |
| 2021 | Arora et al. | Integration of multi-resolution ground segmentation. | Ignores dynamic residues attached to the ground. |
These methods can effectively eliminate ghost trails but require substantial memory and compute resources. They also depend on highly accurate pose estimates, which is a challenging prerequisite for SLAM itself.
3. Visibility-Based Methods
This method operates on a simpler geometric principle: a LiDAR beam travels in a straight line. If two points, \(A\) (near) and \(B\) (far), lie on the same beam (or very close rays) in different scans, the nearer point \(A\) is likely dynamic because it occluded point \(B\) previously. By comparing incoming scans against a maintained map, points in the new scan that are much closer than the expected range stored in the map for that ray direction are flagged as dynamic.
Pomerleau et al. introduced a Bayesian framework incorporating surface normal and incidence angle to handle ambiguity. Removert employed a multi-resolution range image representation to first aggressively remove dynamic points and then selectively recover potentially misclassified static points. RF-LIO integrated an adaptive multi-resolution range image into a LiDAR-Inertial Odometry framework for real-time dynamic removal. BeautyMap proposed a fast binary encoding scheme for global maps and query scans to quickly identify inconsistent regions. Some methods combine visibility with other techniques; for instance, DORF uses a coarse visibility-based removal followed by a fine-grained Bird’s Eye View (BEV) detection of ground dynamic points.
The core visibility check can be formalized. Let \(M\) be the existing map providing an expected range \(r_m(\theta, \phi)\) for a given bearing \((\theta, \phi)\). A new point \(p\) from a scan at that bearing has a measured range \(r_p\). If \(r_p < r_m – \delta\), where \(\delta\) is a tolerance threshold, then \(p\) is considered a dynamic candidate. Formally, the dynamic label \(D(p)\) is:
$$
D(p) =
\begin{cases}
1 & \text{if } r_p < r_m(\theta_p, \phi_p) – \delta \\
0 & \text{otherwise}
\end{cases}
$$
| Year | Author(s) | Key Contribution | Limitation |
|---|---|---|---|
| 2020 | Kim et al. | Removert: Multi-resolution range image static point recovery. | Static points may be erroneously removed initially. |
| 2021 | Qian et al. | RF-LIO: Tightly-coupled LiDAR-Inertial Odometry with dynamic removal. | Performance degrades in feature-sparse, open areas. |
| 2023 | Chen et al. | DORF: Two-stage (coarse-to-fine) removal framework. | Primarily suited for relatively flat ground. |
| 2024 | Jia et al. | BeautyMap: Fast binary matrix comparison for dynamic region detection. | Difficulty handling multi-level ground scenes. |
Visibility-based methods are computationally cheaper than ray-tracing as they don’t maintain a full volumetric map. However, they suffer from two main issues: misclassification of ground points when the incidence angle is near 90 degrees, and failure to detect dynamic points that are fully occluded by larger dynamic objects (e.g., a pedestrian behind a bus).
Processing Strategies for Different Dynamics in SLAM Frameworks
The classical 3D LiDAR SLAM framework comprises sensor data processing, front-end odometry, back-end optimization, loop closure detection, and map building. Dynamic objects interfere primarily with the front-end (corrupting scan matching) and the map-building module (creating artifacts). However, not all non-static objects are equal. For a more nuanced strategy, we classify objects by their dynamic level:
- Highly-Dynamic Objects: Continuously moving (e.g., walking people, driving cars).
- Low-Dynamic Objects: Temporarily stationary (e.g., people standing, cars at a traffic light).
- Semi-Dynamic Objects: Stationary during a SLAM session but movable between sessions (e.g., chairs, parked cars, temporary materials).
- Static Objects: Permanently fixed (e.g., buildings, roads).
Different processing strategies are tailored to handle these categories effectively within the SLAM pipeline.
1. Online Real-Time Processing Strategy
This strategy aims to identify and remove dynamic points in real-time, primarily targeting highly-dynamic objects. It can be implemented in the front-end (before/during scan matching) or the back-end (during map update).
Front-End Implementation: Often integrates with LiDAR-Inertial Odometry (LIO). RF-LIO uses visibility-based removal within a tightly-coupled LIO framework. DRR-LIO employs a voxel occupancy descriptor for removal before scan matching. ID-LIO leverages ground height information and a delayed removal strategy across spatial and temporal dimensions. To address high computational cost, Dynamic-LIO proposes a label-consistency method for very fast (10ms) removal. Beyond filtering, methods like TRLO integrate a 3D object detector (e.g., PointPillars, CenterPoint) and tracker into the odometry front-end to actively identify and remove dynamic objects’ bounding boxes.
Back-End Implementation: Focuses on cleaning the global map. SuMa++ uses semantic segmentation labels to filter dynamic points while building a semantic surfel map. DynamicFilter accelerates the ray tracing process by first creating a static submap using visibility checks and then refining it with a map-to-map optimization backend. RH-Map uses a two-level hashed map structure for efficient online map updates and dynamic object removal.
This strategy is fast but may leave residual dynamic points or misclassify static ones, especially at long ranges, due to sparse and ambiguous data.
| Locus | Author(s) | Key Contribution | Limitation |
|---|---|---|---|
| Front-End | Qian et al. (RF-LIO) | Dynamic removal in tight LIO coupling. | Fails if sensor FOV is fully occluded. |
| Wu et al. (ID-LIO) | Spatio-temporal delayed removal strategy. | Prone to drift in Z-direction in open areas. | |
| Yuan et al. (Dynamic-LIO) | Extremely fast label-consistency method. | Lacks validation in highly complex environments. | |
| Jia et al. (TRLO) | Integration of 3D detection & tracking into odometry. | Does not integrate IMU or loop closure. | |
| Back-End | Chen et al. (SuMa++) | Semantic map representation for filtering. | Fails with a high density of dynamic objects. |
| Fan et al. (DynamicFilter) | Accelerated ray tracing for online map cleaning. | Still suffers from ground point misclassification. |
2. Offline Post-Processing Strategy
This strategy is not constrained by real-time requirements and is typically used to remove both low-dynamic and residual highly-dynamic objects after a SLAM session. It operates on the complete trajectory and all collected point cloud frames or submaps, leveraging global context.
ERASOR (Egocentric Ratio of Pseudo Occupancy-based dynamic object Removal) defines a novel descriptor based on point height differences within a voxel column to segment dynamic points from the static ground, followed by plane fitting for refinement. It assumes a planar ground. Arora et al. improved ground segmentation in OctoMap without a planar assumption for sloped terrain. DORF, as mentioned, combines coarse visibility-based removal with fine BEV-based detection for robust performance in crowded urban settings.
While offline processing achieves more thorough dynamic point removal than online methods by utilizing full sequence information, it is not real-time and may struggle in scenes with an extremely high percentage of dynamic points.
| Year | Author(s) | Key Contribution | Limitation |
|---|---|---|---|
| 2021 | Lim et al. (ERASOR) | Pseudo occupancy ratio descriptor for ground-object separation. | Assumes planar ground, fails on uneven terrain. |
| 2023 | Chen et al. (DORF) | Hybrid coarse (visibility) and fine (BEV) removal. | Designed for flat urban environments. |
3. Lifelong SLAM Strategy
This strategy is essential for handling semi-dynamic objects and long-term operation. It focuses on maintaining and updating the map over multiple SLAM sessions (days, months), detecting changes, and keeping the map current.
Early works like DPG-SLAM used incremental smoothing for pose graph optimization in changing environments. FreMEn modeled environment dynamics using a frequency-based approach to predict future states. Modular frameworks like LT-mapper implement a “remove-then-revert” strategy across sessions and maintain an incremental changelog for map updates. Recent systems, such as the one by Chen et al., focus on continuous map updating via a multi-session approach, using a 2.5D grid map for fast change detection and filtering of a 3D prior map. To address storage concerns, Yang et al. proposed a sensor-agnostic modular framework that reconstructs past session maps on-demand without storing raw data, significantly improving memory efficiency.
Lifelong SLAM is the most comprehensive but also the most complex, involving continuous map management, change detection, and consistency maintenance.
| Year | Author(s) | Key Contribution | Limitation |
|---|---|---|---|
| 2017 | Krajník et al. (FreMEn) | Spectral model for predicting periodic environment changes. | High computational overhead. |
| 2021 | Kim et al. (LT-mapper) | Modular framework for lifelong mapping with change detection. | Maintaining multiple maps affects real-time performance. |
| 2024 | Yang et al. | Memory-efficient, session-reconstructible lifelong framework. | Some static points (1-3%) may still be misclassified. |
Evaluation Metrics and Datasets
1. Evaluation Metrics
Performance is evaluated based on localization accuracy and map quality.
Localization Accuracy:
Absolute Trajectory Error (ATE): Measures the global consistency of the estimated trajectory \(\hat{\mathbf{X}} = \{\hat{\mathbf{x}}_1, …, \hat{\mathbf{x}}_M\}\) against the ground truth \(\mathbf{X} = \{\mathbf{x}_1, …, \mathbf{x}_M\}\). The trajectories are first aligned using a rigid transformation \(\mathbf{S}\). The RMSE of ATE is:
$$ATE_{RMSE} = \sqrt{ \frac{1}{M} \sum_{i=1}^{M} \| \mathbf{x}_i – \mathbf{S}\hat{\mathbf{x}}_i \|^2 }$$
Relative Pose Error (RPE): Measures the local drift over a fixed time interval \(\Delta t\). For poses at times \(i\) and \(i+\Delta t\), the error is:
$$\mathbf{E}_i = (\mathbf{Q}_i^{-1} \mathbf{Q}_{i+\Delta t})^{-1} (\mathbf{P}_i^{-1} \mathbf{P}_{i+\Delta t})$$
where \(\mathbf{P}, \mathbf{Q}\) are estimated and ground truth poses. The translational RPE RMSE is commonly reported.
Map Quality (Dynamic Removal): For a static point cloud map output \(M_{static}\) and a ground truth static map \(M_{gt}\), points are classified.
- True Positive (TP): Static point correctly kept.
- False Positive (FP): Dynamic point incorrectly kept.
- False Negative (FN): Static point incorrectly removed.
- True Negative (TN): Dynamic point correctly removed.
Precision and Recall are defined as:
$$\text{Precision} = \frac{TP}{TP + FP}, \quad \text{Recall} = \frac{TP}{TP + FN}$$
For highly dynamic scenes, Preservation Rate (PR) and Rejection Rate (RR) are more stable:
$$PR = \frac{P_{ss}}{P_{is}} \times 100\%, \quad RR = \left(1 – \frac{P_{id}}{P_{sd}}\right) \times 100\%$$
where \(P_{ss}\): static points in result, \(P_{is}\): static points in input, \(P_{id}\): dynamic points in result, \(P_{sd}\): dynamic points in input.
2. Benchmark Datasets
Several public datasets facilitate the evaluation of dynamic SLAM algorithms for embodied AI robots.
| Dataset | Year | Environment | Key Features for Dynamic SLAM |
|---|---|---|---|
| KITTI Odometry | 2012 | Outdoor (Urban/Rural/Highway) | Standard benchmark with GT trajectories; contains moving cars. |
| NCLT | 2016 | Long-term Indoor & Outdoor | Long-term, multi-session data with seasonal and structural changes. |
| SemanticKITTI | 2019 | Outdoor | Provides dense per-point semantic labels for 28 classes (moving/static). |
| UrbanLoco | 2020 | Outdoor (Dense Urban) | High-density urban driving in San Francisco & Hong Kong. |
| DOALS | 2021 | Indoor (Crowded) | Specifically recorded with many pedestrians; includes manual annotations. |
| Dynablox | 2023 | Indoor & Outdoor | Contains atypical dynamic objects (thrown balls, rolling luggage). |
Future Research Directions
The evolution of dynamic SLAM for embodied AI robots points towards several promising research avenues.
1. Deep Learning Integration: Moving beyond semantic segmentation, future methods could leverage powerful 3D object detectors and multi-object trackers to directly identify and excise dynamic objects before they corrupt the scan-matching process. A significant challenge remains the accurate detection of small dynamic targets (e.g., construction workers) amidst larger objects in cluttered environments like construction sites, which is critical for the safety and efficiency of embodied AI robots.
2. Multi-Sensor Fusion: To operate robustly in complex, unstructured environments, embodied AI robots will increasingly rely on the fusion of LiDAR, cameras, and IMUs. While LiDAR-IMU fusion is common (e.g., LIO-SAM, FAST-LIO2), the integration of camera data for providing rich texture, enabling visual-inertial odometry (VIO), and enhancing loop closure detection (as in LVI-SAM) will be crucial for achieving robustness in diverse and highly dynamic conditions.
3. Lightweight and Scalable Architectures: Many state-of-the-art algorithms are computationally demanding. Developing lightweight models that can run efficiently on the resource-constrained hardware typical of mobile embodied AI robots is essential. Furthermore, scaling these systems to handle lifelong operation over large areas requires innovative map representations (e.g., neural implicit maps, sparse parametric maps) and efficient change detection mechanisms that manage memory and compute growth.
Conclusion
LiDAR SLAM is a cornerstone technology for enabling autonomous navigation in embodied AI robots. The assumption of a static world, however, breaks down in practical applications, necessitating robust dynamic point removal. This review has categorized and analyzed the primary methodologies—semantic segmentation, ray tracing, and visibility-based—alongside the strategic approaches of online processing, offline post-processing, and lifelong SLAM. Each method and strategy presents inherent trade-offs between accuracy, computational cost, generality, and real-time capability. Current challenges include reliance on labeled data, high memory footprints, sensitivity to occlusion, and ground point misclassification. The future of dynamic SLAM for embodied AI robots lies in the deeper integration of learning-based perception, robust multi-sensor fusion, and the development of efficient, scalable systems capable of lifelong environment understanding. Advancements in these areas will be pivotal in deploying reliable and intelligent embodied AI robots in the ever-changing real world.