DEV Community

Richard Minsoo Go
Richard Minsoo Go

Posted on

Autonomous Driving Software — Ch.04: Drivable Area Estimation [EN]

STATERA Library · Series 08 · Volume 12 · Chapter 04 · EN
Source: https://github.com/Statera-Library/08_Vol_12_Autonomous_Driving_Software/blob/main/Chapter_04_Drivable_Area_Estimation/Chapter_04_Drivable_Area_Estimation_EN.md

Volume 12. Autonomous Driving Software

Chapter 04. Drivable Area Estimation

04.01. Drivable Area Estimation Methods Overview

Drivable area estimation determines which portions of the surrounding environment can be safely occupied by an autonomous vehicle or mobile robot. Unlike object detection, which identifies discrete entities such as pedestrians or vehicles, drivable area estimation reasons about continuous space. Its output forms a geometric and semantic representation of where the platform may move while respecting terrain, obstacles, boundaries, and operational constraints.

Within an autonomous driving software stack, drivable area estimation connects sensor processing with planning and control. Raw camera images, LiDAR point clouds, radar observations, localization information, and map data are transformed into representations such as free-space polygons, segmentation masks, occupancy grids, or traversability maps. These representations allow downstream planners to distinguish feasible motion regions from occupied, uncertain, or prohibited space.

A fundamental distinction exists between free-space detection and traversability estimation. Free space indicates that no obstacle has currently been observed in a region, but this does not necessarily mean that the region can be traversed safely. A surface may appear unoccupied while containing excessive slope, loose soil, deep water, vegetation, stairs, potholes, or other terrain properties that exceed the vehicle\'s physical capabilities.

Classical geometric approaches estimate drivable space primarily from sensor geometry. LiDAR-based systems analyze point heights, local surface normals, gradients, discontinuities, and point density to separate ground from obstacles. After ground segmentation, regions without detected obstacles can be projected onto a two-dimensional plane or bird\'s-eye-view representation. These methods are computationally efficient and remain valuable when explicit geometric reasoning is required.

Camera-based approaches infer drivable regions from visual appearance and scene semantics. Traditional techniques used color, texture, edges, vanishing points, and geometric assumptions, while modern systems typically employ convolutional neural networks or transformer-based semantic segmentation models. Pixels may be classified into road, sidewalk, grass, vegetation, obstacle, unknown, and other categories before being transformed into a spatial representation suitable for navigation.

Bird\'s-eye-view representations are particularly useful because motion planning operates naturally in a ground-referenced coordinate system. Perspective camera observations can be projected into BEV using geometric transformations or learned view transformations. LiDAR and radar measurements can also be represented in the same coordinate frame, enabling multiple sensor modalities to contribute to a unified description of the space surrounding the vehicle.

Occupancy grids provide another widely used representation. The environment is divided into cells whose states indicate free, occupied, or unknown space, often with probabilistic values rather than binary labels. Measurements accumulated over time update these probabilities as the robot moves. Occupancy representations are especially useful for local planning because they provide a compact interface between perception and collision-aware trajectory generation.

Semantic occupancy extends geometric occupancy by describing what occupies each region. Instead of merely determining that a cell is occupied, the system may distinguish vehicles, pedestrians, vegetation, curbs, walls, terrain, or other semantic classes. This additional information enables planners to apply different policies to different environmental structures and provides a foundation for richer scene understanding and predictive autonomous behavior.

Multi-sensor fusion improves robustness by exploiting complementary sensing characteristics. Cameras provide dense appearance and semantic information, LiDAR provides accurate three-dimensional geometry, radar contributes velocity and robust ranging under difficult visibility conditions, while GNSS and inertial localization place observations in consistent coordinates. Fusion can occur at raw-data, feature, object, occupancy, or decision levels depending on system architecture and computational constraints.

Outdoor AMRs require a broader interpretation of drivable space than conventional road vehicles. Their operational domains may include sidewalks, industrial yards, construction sites, ports, campuses, gravel roads, grass, ramps, and partially structured terrain. Consequently, lane boundaries alone cannot define mobility. The system must estimate whether terrain is physically traversable according to vehicle dimensions, ground clearance, wheel configuration, slope limits, traction, and stability.

Traversability can therefore be represented as a continuous cost rather than a binary decision. Flat pavement may receive a low traversal cost, while rough ground, steep slopes, narrow passages, or uncertain terrain receive progressively higher costs. Completely blocked or physically infeasible regions receive prohibitive costs. Such cost maps allow planners to optimize trajectories by balancing distance, safety margin, terrain difficulty, energy consumption, and mission objectives.

Negative obstacles create a particularly difficult perception problem. A wall or parked vehicle produces positive geometric returns, whereas a ditch, hole, downward step, or missing road surface may be characterized by an absence or sudden disappearance of measurements. Detecting these hazards requires reasoning about surface continuity, expected ground geometry, visibility, sensor occlusion, and temporal evidence rather than simply searching for elevated points.

Dynamic environments further require the drivable-area representation to evolve continuously. A region that was free seconds earlier may become occupied by a pedestrian, vehicle, forklift, or another robot. Conversely, temporarily blocked space may become available again. Sensor observations must therefore be time-aligned and repeatedly integrated so that the representation reflects current conditions rather than preserving stale obstacles indefinitely.

Uncertainty is essential because absence of evidence must not automatically be interpreted as evidence of free space. Occluded regions, areas beyond sensor range, surfaces with poor reflectivity, severe shadows, precipitation, dust, or sparse point-cloud coverage may remain unknown. A safety-oriented system explicitly represents this uncertainty and allows planning policies to treat unknown regions differently from confidently observed free regions.

Map information can complement online perception by providing prior knowledge about roads, lanes, sidewalks, restricted zones, buildings, and expected boundaries. However, maps should not replace real-time sensing because construction, parked equipment, temporary barriers, vegetation, and environmental changes can invalidate prior information. Practical systems therefore combine map priors with current observations and preserve mechanisms for detecting disagreements between them.

The spatial resolution and temporal update rate of drivable-area estimation directly affect planning quality and computational demand. Fine grids preserve narrow obstacles and detailed boundaries but consume more memory and processing time. Coarse representations are efficient but may hide small hazards. Designers therefore select resolution, sensing range, update frequency, and representation format according to vehicle speed, stopping distance, platform dimensions, and available compute resources.

Drivable area estimation must also account for the physical footprint of the vehicle. A point may lie in free space while the complete vehicle body cannot pass through the surrounding geometry. Obstacle inflation, footprint convolution, swept-volume analysis, or configuration-space reasoning converts perceived environmental geometry into regions that are feasible for a platform with finite width, length, turning radius, and steering constraints.

Modern learning-based methods increasingly estimate geometry, semantics, depth, occupancy, and traversability jointly. Multi-task networks can share representations across segmentation, depth estimation, object detection, and BEV prediction, reducing redundant computation and improving contextual reasoning. Learned approaches can recognize complex terrain patterns that are difficult to encode manually, but their reliability depends strongly on training coverage, domain diversity, calibration, and uncertainty handling.

Hybrid architectures combine learned perception with explicit geometric reasoning. A neural network may identify semantic terrain classes while LiDAR estimates surface geometry and an analytical vehicle model determines whether the resulting terrain is traversable. This separation is useful because semantic recognition and physical feasibility are different questions. A surface classified as grass, for example, may be traversable for one AMR configuration but prohibited for another.

The resulting drivable-area representation ultimately serves as a contract between perception and motion planning. It should communicate not only where the robot can move, but also confidence, terrain cost, obstacle boundaries, semantic restrictions, and relevant temporal changes. The attached software structure consequently develops this topic from LiDAR and camera methods through sensor fusion, occupancy grids, off-road traversability, negative obstacles, dynamic updates, validation, and an outdoor AMR case study.

For Physical AI systems, drivable area estimation can evolve beyond static perception toward predictive environmental understanding. A world model may estimate how free space will change as agents move, how terrain properties affect future vehicle motion, and how uncertainty propagates through planned actions. This transforms drivable-area estimation from simple road segmentation into an active spatial reasoning process connecting perception, vehicle embodiment, prediction, planning, and safe autonomous action.

04.02. LiDAR Based Free Space Detection [w/Code]

LiDAR-based free-space detection identifies regions around an autonomous vehicle or AMR where no physical obstacle is detected and where movement may potentially be possible. Unlike camera segmentation, which primarily reasons from visual appearance, LiDAR provides direct three-dimensional geometric measurements. The fundamental task is therefore to transform a point cloud into a spatial representation that distinguishes ground or traversable surfaces from obstacles, unknown regions, and geometrically unsafe areas.

The processing pipeline normally begins with LiDAR point-cloud acquisition and preprocessing. Raw measurements may contain noise, outliers, reflections, motion distortion, and points originating from the robot itself. Filtering, coordinate transformation, and downsampling are therefore applied before free-space estimation. The processed points are transformed into a vehicle-centered coordinate frame so that subsequent ground estimation, obstacle detection, and spatial projection can be performed consistently with localization and motion-planning modules.

Ground estimation is one of the central operations in LiDAR-based free-space detection. A typical approach analyzes the height and local geometric structure of points relative to the expected ground surface. Methods based on height thresholds, local slope, surface normals, plane fitting, elevation maps, or scan-line characteristics can separate probable ground points from elevated obstacles. For outdoor AMRs, however, the ground cannot always be modeled as a single flat plane because roads, ramps, slopes, gravel, and uneven terrain can produce significant elevation changes.

After ground estimation, the remaining points are interpreted as potential obstacles or non-traversable structures. Objects such as vehicles, pedestrians, walls, poles, curbs, barriers, and equipment produce geometric discontinuities above the estimated ground surface. The system can project these measurements into a two-dimensional bird\'s-eye-view representation and mark corresponding regions as occupied. This representation provides a direct spatial interface for local planning because obstacle locations are expressed relative to the robot rather than only as image coordinates.

Free space can then be estimated from the relationship between observed ground and occupied regions. Areas containing valid ground observations and without detected obstacles are considered candidate free space, while regions containing obstacles are classified as occupied. Areas outside the LiDAR observation range or hidden behind obstacles should generally remain unknown rather than automatically being classified as free. This distinction is important because LiDAR measures only the portions of the environment that are geometrically visible to the sensor.

Ray-based reasoning can further improve free-space estimation. A LiDAR beam travels from the sensor toward a measured return, and the space before the return provides evidence that the beam path was not occupied by a detected obstacle. The measured endpoint can represent an obstacle or surface boundary depending on its height and local geometry. By accumulating these ray observations into a grid, the system can estimate free, occupied, and unknown cells while preserving the relationship between sensor visibility and environmental geometry.

The resulting representation is often converted into an occupancy grid or local cost map. Each cell can store a state or probability describing whether the region is free, occupied, or unknown. Temporal integration allows multiple LiDAR scans to reinforce stable observations and reduce the effect of individual noisy measurements. However, temporal accumulation must also account for moving objects and changing environments. Stale obstacle information can otherwise remain in the map and incorrectly restrict subsequent motion planning.

Vehicle geometry must be considered after free-space extraction. A LiDAR point may indicate that a particular location is free while the complete AMR footprint cannot pass through that region. The free-space representation therefore needs to account for vehicle width, length, ground clearance, turning characteristics, and safety margins. Obstacle inflation or footprint-based spatial filtering can convert point-level observations into a representation that more accurately reflects whether the entire platform can physically occupy a region.

Outdoor AMRs introduce additional challenges because their environments are not necessarily structured around conventional road geometry. A robot may operate on pavement, grass, gravel, soil, ramps, industrial yards, construction areas, or mixed terrain. LiDAR geometry can identify whether a surface is physically present, but free space alone does not guarantee traversability. Terrain slope, roughness, step height, clearance, expected traction, and vehicle stability may need to be evaluated separately before a candidate region is accepted as operationally traversable.

Negative obstacles require particular attention because they may not generate conventional obstacle returns. Pits, trenches, drop-offs, holes, and missing ground can appear as discontinuities or unexpected gaps in the point cloud. A simple rule that treats every unoccupied region as free can therefore create dangerous decisions. Reliable detection requires reasoning about expected ground continuity, elevation changes, point distribution, sensor visibility, and the geometry of the surrounding surface to distinguish genuine free space from potentially hazardous unknown space.

LiDAR-based free-space detection is also sensitive to sensor configuration and environmental conditions. Sensor height, vertical field of view, angular resolution, range, mounting position, and scan frequency determine which portions of the environment can be observed. Dust, rain, fog, vegetation, reflective materials, transparent surfaces, and sparse returns can degrade geometric reliability. Consequently, a practical system should associate confidence or uncertainty with its free-space estimate rather than treating every LiDAR-derived cell as equally reliable.

For an autonomous driving software architecture, LiDAR-based free-space detection is best regarded as a perception component that supplies structured spatial evidence to downstream modules. Its output can be fused with camera-based semantic information, radar observations, localization, and map data to construct a more complete drivable-area representation. This layered approach separates geometric evidence from semantic interpretation while allowing the final planning system to reason about obstacles, free space, uncertainty, and vehicle-specific traversability within a common spatial framework.

04.03. Camera Based Road Segmentation BEV [w/Code]

Camera-based road segmentation estimates which image regions correspond to road or potentially drivable surfaces by analyzing visual information captured from onboard cameras. Unlike LiDAR-based methods, which rely primarily on three-dimensional geometry, camera segmentation uses appearance, texture, color, shape, context, and learned semantic representations. Its output can provide dense spatial information that is particularly useful for identifying road boundaries, sidewalks, vegetation, vehicles, and other scene elements before transforming the result into a representation suitable for autonomous navigation.

The processing pipeline normally begins with camera image acquisition and preprocessing. Lens distortion, exposure variation, image noise, motion blur, and differences between cameras can affect segmentation quality. Image rectification, normalization, resizing, and color-space processing may therefore be applied before inference. The camera coordinate system must also be calibrated with respect to the vehicle coordinate system so that image-based road regions can later be transformed into a geometrically meaningful representation for planning.

Traditional road segmentation methods relied on manually designed visual features. Color thresholds, edge detection, texture analysis, region growing, vanishing-point estimation, and perspective geometry were commonly used to distinguish road surfaces from surrounding environments. These methods can be computationally efficient and understandable, but their performance can degrade when road appearance changes because of lighting, weather, pavement materials, shadows, vegetation, or unusual road structures.

Modern systems generally use deep neural networks for semantic segmentation. Convolutional neural networks can extract hierarchical spatial features from images, while encoder-decoder architectures preserve both high-level semantic information and fine spatial details. More recent transformer-based models can use broader contextual relationships across the image. The segmentation network may classify each pixel into categories such as road, sidewalk, grass, vegetation, obstacle, building, or unknown, creating a dense semantic representation of the visible environment.

Road segmentation should not be interpreted as a simple binary classification problem in an autonomous robot. A visually road-like region may contain temporary obstacles, damaged pavement, water, snow, construction materials, or surfaces that are unsuitable for the vehicle. Conversely, an AMR may intentionally operate outside conventional roads on grass, gravel, industrial yards, or other terrain. The semantic segmentation result should therefore provide environmental evidence that is subsequently combined with geometry and vehicle-specific traversability reasoning.

A major advantage of camera-based segmentation is its ability to capture semantic boundaries that may be difficult to obtain from sparse geometric measurements. Curbs, painted road boundaries, pedestrian areas, vegetation, and visually distinctive terrain can often be recognized through appearance and context. This semantic information can complement LiDAR geometry by explaining what a surface represents rather than only describing its physical shape.

BEV transformation converts image-based segmentation from perspective coordinates into a ground-referenced representation. A conventional image contains strong perspective distortion, so identical physical distances appear differently depending on their position in the image. Camera calibration, camera pose, depth estimation, inverse perspective mapping, or learned view transformation can be used to generate a bird\'s-eye-view representation in which road regions and surrounding structures are expressed relative to the vehicle.

The quality of BEV road segmentation depends strongly on geometric calibration and depth reasoning. A simple perspective transformation may work adequately when the ground is approximately planar, but it becomes less reliable on slopes, uneven terrain, ramps, or irregular outdoor environments. Learned BEV approaches can infer spatial structure directly from visual features, while explicit depth or multi-view information can provide additional geometric constraints. The selected method should reflect the expected operational environment and required accuracy.

Camera-based segmentation also benefits from temporal information. Consecutive frames contain overlapping observations of the same road and environmental structures, allowing the system to stabilize segmentation boundaries and reduce frame-to-frame fluctuations. Temporal filtering, optical-flow information, feature tracking, or sequential neural models can help maintain a consistent road representation. However, moving objects and sudden environmental changes must be handled carefully so that temporal smoothing does not preserve obsolete information.

Outdoor operation introduces substantial variation in visual conditions. Illumination changes between day and night, shadows can resemble boundaries, rain and fog reduce visibility, and wet surfaces can change road appearance. Seasonal vegetation, dust, snow, construction activity, and different pavement materials can also produce domain shifts. Robust camera segmentation therefore requires diverse training data, appropriate augmentation, calibration, confidence estimation, and validation across the expected operational design domain.

Camera segmentation becomes significantly more useful when fused with other perception sources. LiDAR can provide precise surface geometry and obstacle locations, while cameras provide dense semantic classification. Radar can contribute motion and range information, and localization and map data can provide spatial context. A fusion system can therefore combine the semantic evidence from camera segmentation with geometric and temporal evidence from other sensors to construct a more reliable drivable-area representation.

For an outdoor AMR, the final output should not simply be a road mask. It should provide information that can be transformed into a drivable-area or traversability representation while considering vehicle footprint, terrain properties, obstacles, uncertainty, and operational restrictions. Camera-based road segmentation therefore serves as an important perception layer between raw visual sensing and spatial reasoning. When integrated with BEV processing and multi-sensor fusion, it provides a dense semantic foundation for downstream planning, trajectory generation, and safe autonomous navigation.

04.04. Multi Sensor Drivable Area Fusion [w/Code]

Multi-sensor drivable-area fusion combines complementary observations from cameras, LiDAR, radar, localization systems, and map information to construct a more reliable representation of the space available for autonomous motion. Each sensor observes the environment differently: cameras provide dense semantic information, LiDAR provides accurate three-dimensional geometry, radar provides range and motion information under challenging visibility, while localization and maps provide spatial reference. Fusion therefore reduces dependence on any single sensing modality.

The primary objective is not simply to combine sensor outputs, but to establish a common spatial representation in which heterogeneous measurements can be compared and integrated. Camera segmentation may describe a region as road or grass, while LiDAR may indicate its surface elevation and obstacle geometry. Radar may identify a moving object that is difficult to segment visually. These observations must be transformed into a consistent coordinate system and associated with appropriate timestamps before they can contribute reliably to a shared drivable-area representation.

Sensor calibration is a fundamental requirement for reliable fusion. Camera intrinsic parameters describe the internal optical characteristics, while extrinsic calibration defines the spatial relationship between the camera, LiDAR, radar, and vehicle coordinate frames. Even small calibration errors can cause projected points or semantic regions to shift relative to one another. For an outdoor AMR, calibration should therefore be treated as an operational engineering parameter that requires verification after sensor replacement, mechanical modification, or significant vibration.

Time synchronization is equally important because the environment is dynamic. A camera image, LiDAR scan, and radar measurement acquired at slightly different times may correspond to different positions of a moving pedestrian or vehicle. Hardware timestamps, synchronized clocks, interpolation, and motion compensation can reduce these inconsistencies. The fusion pipeline should preserve temporal information so that an observation is not treated as spatially correct when it has already become outdated.

Fusion can be implemented at several abstraction levels. Raw-data fusion combines sensor measurements before high-level interpretation, while feature-level fusion combines learned or geometric features extracted from different modalities. Object-level fusion associates independently detected objects, and occupancy-level fusion integrates spatial evidence directly into a common grid or BEV representation. Decision-level fusion combines the outputs of separate perception systems. The appropriate level depends on sensor characteristics, computational resources, latency requirements, and system architecture.

A common practical representation is a local BEV or occupancy grid in the vehicle coordinate frame. Camera-based semantic segmentation can be projected into this representation, while LiDAR points are accumulated according to their geometric positions and radar detections are inserted according to range and velocity. Each grid cell can maintain multiple attributes, such as occupancy probability, semantic class, traversability cost, confidence, and observation time. This creates a richer spatial representation than any individual sensor can normally provide.

Sensor fusion should also distinguish between positive evidence and missing evidence. A LiDAR return can provide strong evidence that a physical surface or obstacle exists, while the absence of a return does not necessarily prove that the corresponding region is free. Camera visibility may be blocked by an object, radar may detect through some visual conditions, and map information may provide prior knowledge where current sensing is incomplete. A fusion system should therefore represent free, occupied, and unknown states together with confidence rather than forcing every region into a binary classification.

Semantic and geometric information provide different but complementary evidence. A camera may recognize vegetation, road markings, pedestrians, or vehicles even when their geometric structure is difficult to distinguish. LiDAR can determine surface height, slope, obstacle boundaries, and three-dimensional structure even when visual appearance is ambiguous. Combining these observations allows the system to distinguish visually plausible but physically unsuitable areas from genuinely traversable surfaces. This is particularly important for outdoor AMRs operating beyond conventional road environments.

Radar contributes another dimension by measuring range and relative motion. It can remain useful when lighting is poor or visibility is degraded by rain, fog, or other environmental conditions. Radar observations can therefore support the identification of dynamic obstacles and provide temporal information that complements camera and LiDAR perception. Because radar measurements can be sparse or ambiguous, they should generally be integrated with other evidence rather than interpreted independently as a complete drivable-area solution.

Map information and localization can provide additional spatial priors during fusion. A map may describe expected roads, paths, restricted zones, or environmental boundaries, while localization determines where the current sensor observations belong within that spatial context. However, the map represents prior information and may differ from the current environment because of construction, temporary obstacles, parked equipment, vegetation, or other changes. Online sensor observations should therefore be capable of overriding outdated map assumptions when sufficient evidence indicates that the environment has changed.

The final fused representation should be converted into a vehicle-aware drivable-area or traversability model. Vehicle dimensions, ground clearance, turning radius, slope capability, traction, stability, and safety margins determine whether an apparently free region can actually be occupied by the platform. Terrain information can further assign different traversal costs to pavement, grass, gravel, mud, slopes, or rough surfaces. Consequently, fusion provides the evidence, while vehicle-specific reasoning determines how that evidence should influence autonomous motion.

A robust multi-sensor fusion architecture should continuously evaluate sensor health and disagreement. A failed camera, degraded LiDAR return pattern, radar malfunction, inaccurate localization estimate, or inconsistent map association should not silently produce a normal-looking output. Confidence monitoring, sensor-status information, temporal consistency checks, and fallback behavior can identify degraded conditions and allow the system to reduce reliance on unreliable observations. The resulting architecture provides a structured connection between heterogeneous perception, drivable-area estimation, traversability reasoning, and downstream planning for outdoor autonomous robots.

04.05. Occupancy Grid Map Construction and Update [w/Code]

An occupancy grid map represents the surrounding environment as a regular two-dimensional or three-dimensional grid in which each cell stores an estimate of whether the corresponding region is free, occupied, or unknown. For autonomous driving and outdoor AMR systems, this representation provides a practical interface between perception and motion planning. Instead of passing raw LiDAR points or camera pixels directly to the planner, the perception system converts heterogeneous observations into a spatial map that can be continuously queried for collision avoidance and navigation.

The construction process begins by defining the grid coordinate system, resolution, spatial range, and reference frame. A vehicle-centered local grid is commonly useful because it can be updated efficiently as the robot moves. The resolution must balance geometric detail against computational and memory requirements. A fine grid can preserve small obstacles and narrow passages, whereas a coarse grid reduces processing cost but may remove important spatial details. The selected resolution should therefore reflect vehicle dimensions, operating speed, sensor range, and planning requirements.

Sensor observations are transformed into the grid coordinate system before updating individual cells. LiDAR points can be projected according to their three-dimensional positions, while camera-based segmentation can contribute semantic information and radar can provide additional detections and motion evidence. Localization information determines the relationship between the current vehicle pose and the map coordinate system. When multiple sensors are used, calibration and timestamp consistency are essential because spatial or temporal errors can place observations in incorrect cells and produce unstable occupancy estimates.

A basic occupancy grid distinguishes free, occupied, and unknown states. A cell becomes occupied when sensor evidence indicates that an obstacle or physical structure exists within the corresponding region. Free-space evidence can be generated from observed ground surfaces or from the portion of a sensor ray between the sensor origin and a measured obstacle. Unknown cells represent areas that have not been sufficiently observed, such as regions outside the sensor range or areas hidden behind obstacles. Maintaining the unknown state prevents the system from incorrectly treating unobserved space as guaranteed free space.

Probabilistic occupancy mapping provides a more flexible representation than a simple binary grid. Instead of assigning an absolute state immediately, each cell maintains an occupancy probability that is updated as new observations arrive. Bayesian or log-odds formulations can accumulate repeated evidence efficiently. A strong obstacle observation increases the estimated occupancy, while consistent free-space observations reduce it. This allows the map to represent confidence and tolerate individual noisy measurements rather than allowing a single sensor return to determine the permanent state of a cell.

Ray tracing is an important mechanism for generating free-space evidence from range sensors. For each LiDAR measurement, the system considers the path from the sensor origin toward the measured endpoint. Cells along the visible path can receive free-space evidence, while the endpoint may be marked as occupied when its geometry indicates an obstacle. This process captures the visibility structure of the sensor and prevents the map from being filled with unexplained free cells. Care is required near reflective surfaces, sparse returns, and measurement boundaries where the interpretation of a ray may be uncertain.

Map updating must account for the continuous movement of the robot. A local occupancy grid is typically transformed or reconstructed relative to the current vehicle pose so that the planner always receives a current spatial representation. Efficient implementations may use rolling grids, coordinate transformations, spatial hashing, or other memory-management techniques. The update cycle must also fit within the perception and planning latency budget because an occupancy map that is geometrically accurate but significantly delayed can still produce unsafe or inefficient navigation decisions.

Dynamic objects create another important update problem. A pedestrian, vehicle, forklift, or another AMR can occupy a cell temporarily and then move away. If occupied cells are never cleared, the map gradually becomes contaminated by stale information. Temporal decay, clearing rays, observation timestamps, object tracking, and repeated free-space measurements can remove obsolete occupancy evidence. The update policy should therefore distinguish persistent structures from temporary objects while avoiding excessive clearing that could erase real obstacles during short periods of sensor occlusion.

Semantic information can extend the occupancy grid beyond simple spatial occupancy. A cell may store semantic class, confidence, velocity, terrain type, or traversability cost in addition to occupancy probability. For example, a vehicle and a grass region may both occupy physical cells differently, but the planner should treat them according to different rules. Similarly, a free cell containing steep terrain may not be operationally equivalent to a free cell representing flat pavement. A multi-layer grid can therefore connect geometric occupancy with semantic and physical reasoning.

Outdoor AMRs require particular attention to terrain and negative obstacles. Uneven ground, slopes, gravel, vegetation, curbs, ditches, potholes, and drop-offs may not fit a simple occupied-versus-free model. Elevation maps or terrain layers can be associated with the occupancy grid to estimate surface slope, roughness, and discontinuities. A region with insufficient geometric evidence can remain unknown rather than being declared free. This allows downstream traversability analysis to incorporate both environmental geometry and the physical capabilities of the AMR.

The occupancy grid should ultimately serve as a planning-oriented representation rather than an isolated perception product. Obstacle inflation can incorporate the vehicle footprint and safety margin, while cost layers can represent terrain difficulty, proximity to obstacles, restricted areas, or other navigation preferences. The resulting local map can then be consumed by trajectory generation and motion-planning modules. A well-designed interface keeps the representation sufficiently detailed for collision checking while avoiding unnecessary information that increases computational load.

Validation of occupancy-map construction should examine both spatial accuracy and temporal behavior. Important characteristics include obstacle localization error, free-space accuracy, unknown-area handling, update latency, map persistence, clearing behavior, and robustness to sensor noise. Testing should include static and dynamic scenes as well as difficult outdoor conditions. For the autonomous driving software architecture, occupancy-grid construction therefore acts as a central spatial layer that converts continuously changing multi-sensor observations into an actionable representation for drivable-area estimation, traversability analysis, local planning, and safe autonomous navigation.

04.06. Terrain Traversability Analysis for Off Road AMR [w/Code]

Terrain traversability analysis determines whether an off-road autonomous mobile robot can safely and efficiently move across natural or unstructured terrain. Unlike conventional drivable-area estimation, which often distinguishes road from non-road regions, traversability analysis evaluates the physical interaction between the robot and the terrain. The same surface may be traversable for a lightweight four-wheel AMR but unsuitable for a heavy platform with limited ground clearance, traction, or stability.

The analysis begins with perception of terrain geometry and surface characteristics. LiDAR point clouds, stereo or monocular cameras, depth sensors, radar, and inertial measurements can provide complementary observations of the surrounding ground. These measurements are transformed into a vehicle-centered spatial representation such as an elevation map, voxel map, point-cloud surface, or multi-layer grid. The representation should preserve information relevant to mobility rather than merely indicating whether a region contains an obstacle.

Elevation is one of the most fundamental terrain properties. A local elevation map records estimated ground height across spatial cells and enables the system to identify slopes, steps, depressions, ridges, and discontinuities. Elevation differences between neighboring cells can reveal terrain that exceeds the robot\'s climbing or descending capability. However, elevation measurements are affected by sensor noise, vegetation, occlusion, and sparse observations, so confidence and uncertainty should accompany the geometric estimate.

Terrain slope directly influences mobility, traction, energy consumption, and rollover risk. Longitudinal slope affects climbing and braking requirements, while lateral slope can reduce stability by shifting the effective center of gravity relative to the support polygon. Local surface normals or elevation gradients can be used to estimate slope direction and magnitude. These values are compared with vehicle-specific limits rather than applying a universal threshold, because allowable terrain depends strongly on chassis geometry and loading conditions.

Surface roughness describes smaller-scale height variations that may cause vibration, wheel impact, loss of contact, or reduced controllability. Roughness can be estimated from local point distributions, residual errors from fitted surfaces, height variance, or geometric descriptors derived from neighboring cells. A region may have an acceptable average slope but still be difficult to traverse because of rocks, roots, broken pavement, or repeated surface irregularities that exceed suspension and wheel capabilities.

Step height and ground clearance provide additional geometric constraints. Rocks, curbs, logs, ledges, and abrupt terrain transitions can interfere with wheels or the underbody even when sufficient open space exists around the robot. Traversability analysis should therefore compare detected geometric features with wheel radius, suspension travel, approach angle, departure angle, breakover characteristics, and minimum ground clearance. This converts environmental geometry into platform-specific mobility constraints.

Terrain semantics provide information that geometry alone cannot fully describe. Camera-based perception may classify pavement, gravel, grass, soil, mud, sand, vegetation, water, or other surface categories. Two regions with similar geometric profiles may have very different traction and sinkage properties. Semantic classification can therefore contribute prior knowledge about expected terrain behavior, although the predicted class should not be treated as direct proof of physical traversability.

Traction is especially important for off-road AMRs because wheel-terrain interaction changes substantially across different surfaces. Loose gravel, mud, wet grass, sand, snow, or steep soil may cause longitudinal or lateral slip even when the geometric slope is acceptable. Wheel speed, vehicle velocity, IMU measurements, motor current, and control errors can provide online evidence of slip. Such proprioceptive information allows the robot to refine terrain costs based on its actual physical interaction with the environment.

A traversability map typically combines several terrain attributes into a continuous cost representation. Flat and firm terrain may receive a low traversal cost, while increasing slope, roughness, uncertainty, vegetation, or expected slip raises the cost. Regions that violate hard vehicle limits can be marked non-traversable. This continuous representation is more useful than a simple binary map because the planner can compare alternative routes and trade path length against terrain difficulty, risk, energy consumption, and mission objectives.

Vehicle footprint must be evaluated over the terrain rather than at a single point. A large AMR may simultaneously place its wheels on surfaces with different elevations, producing chassis roll, pitch, or insufficient clearance. Footprint-based terrain evaluation examines the geometry underneath and around the complete vehicle configuration. For articulated, skid-steer, four-wheel-drive, or six-wheel platforms, the evaluation may also account for axle arrangement, wheel contact locations, suspension behavior, and steering constraints.

Negative obstacles such as ditches, holes, trenches, and drop-offs are particularly hazardous in off-road operation. They may generate few direct sensor returns because the surface disappears away from the sensor. Traversability analysis must therefore reason about missing ground, sudden elevation discontinuities, visibility, expected surface continuity, and neighboring observations. Unknown regions near potential drop-offs should generally receive conservative costs until sufficient evidence confirms that they can be traversed safely.

Uncertainty should remain an explicit component of the traversability model. Sparse point clouds, distant terrain, dense vegetation, dust, rain, darkness, and sensor occlusion can reduce confidence in surface estimates. Instead of forcing uncertain cells into traversable or non-traversable categories, the system can maintain confidence values or uncertainty costs. The planner can then prefer well-observed terrain while still allowing uncertain regions to be considered when mission requirements and safety policies permit.

Traversability should also be updated from experience as the robot moves. Predicted terrain properties can be compared with measured wheel slip, body acceleration, suspension response, motor torque, pitch, roll, and achieved velocity. A region predicted to be easy but producing severe slip or vibration indicates that the perception-based terrain model was incomplete. This feedback enables adaptive traversability estimation in which exteroceptive sensing and proprioceptive vehicle response jointly characterize terrain.

For local planning, the traversability map can be combined with obstacle occupancy, dynamic-object information, restricted regions, and route objectives. The planner searches for trajectories that remain collision-free while avoiding terrain that exceeds physical limits or carries excessive cost. At higher speeds, the analysis should consider stopping distance and future terrain along the predicted trajectory rather than evaluating only the area immediately beneath the vehicle. This links terrain perception directly with motion feasibility.

Validation should cover representative off-road conditions rather than only visually diverse scenes. Test environments should include different slopes, rough surfaces, loose materials, vegetation, steps, depressions, wet or degraded terrain, and transitions between surface types. Evaluation should compare predicted traversability with actual vehicle behavior, including successful passage, slip, stability, clearance, speed reduction, intervention, or failure. Such testing connects perception accuracy with the physical performance that ultimately defines traversability.

For an outdoor AMR, terrain traversability analysis therefore acts as the bridge between environmental perception and embodied mobility. It transforms geometric, semantic, inertial, and vehicle-response information into an estimate of where the specific platform can move and at what cost. When integrated with occupancy mapping, vehicle constraints, uncertainty modeling, and local planning, traversability analysis enables autonomous navigation beyond structured roads and supports operation across complex off-road environments.

04.07. Negative Obstacle Detection Pits Curbs [w/Code]

Negative obstacle detection identifies terrain hazards that extend downward from the expected ground surface rather than upward into the robot\'s path. Pits, ditches, trenches, drop-offs, potholes, drainage channels, and descending curbs may produce little or no direct obstacle return from a forward-looking sensor. For an outdoor AMR, these hazards can be more difficult than conventional positive obstacles because the dangerous region may appear as empty space instead of a clearly measurable object.

A positive obstacle such as a wall, rock, vehicle, or raised curb normally generates sensor measurements above the estimated ground plane. A negative obstacle behaves differently because the expected ground surface disappears, drops abruptly, or becomes invisible because of sensor geometry. Detection must therefore reason about missing ground, elevation discontinuities, visibility, and expected surface continuity rather than relying only on the presence of elevated points.

LiDAR provides important geometric evidence for negative obstacle detection. Consecutive laser returns can be analyzed for sudden changes in range, height, local slope, or point density. If beams that previously intersected a continuous ground surface suddenly travel significantly farther before returning, the system may infer a depression or drop-off. However, missing LiDAR returns can also result from low reflectivity, excessive range, water, vegetation, or occlusion, so absence of points alone is insufficient evidence.

Elevation maps provide a useful representation for identifying pits and abrupt terrain transitions. Ground measurements are accumulated into spatial cells, and neighboring elevations are compared to estimate local height differences, slopes, and discontinuities. A sudden downward transition can indicate a ditch or step-down, while a localized depression may represent a pothole or pit. Confidence values should accompany these estimates because sparse measurements can create apparent elevation changes that do not correspond to real terrain hazards.

Curbs require both positive and negative geometric reasoning. An ascending curb produces a raised edge that may interfere with wheels or the vehicle underbody, whereas a descending curb creates a downward step whose safety depends on curb height, wheel radius, suspension travel, ground clearance, and vehicle stability. A curb that is easily traversable by one AMR may exceed the physical capability of another, so detection and traversability assessment should remain separate but closely connected processes.

Camera information can complement geometric sensing by identifying visual cues associated with terrain boundaries. Changes in texture, shading, road edges, curb appearance, drainage structures, and semantic classes may indicate a potential drop or depression before sufficient LiDAR geometry is available. Stereo cameras can additionally estimate depth discontinuities. However, shadows, illumination changes, wet surfaces, vegetation, and visually uniform terrain can resemble negative obstacles, making camera-only decisions unreliable in difficult outdoor conditions.

Sensor fusion improves robustness by combining geometric, semantic, and temporal evidence. LiDAR can provide surface geometry, cameras can recognize curb and terrain context, radar may contribute information about surrounding structures, and localization allows observations to be accumulated consistently over time. A suspected negative obstacle supported by several independent observations can receive greater confidence than a region identified only because one sensor temporarily failed to observe the ground.

Visibility analysis is essential because an unobserved region should not automatically be classified as a pit. Terrain may disappear from the sensor because it lies behind a positive obstacle, outside the field of view, beyond effective range, or at an unfavorable incidence angle. Ray-based reasoning and sensor models can distinguish expected occlusion from unexpected loss of ground returns. Regions that remain ambiguous should be represented as unknown rather than incorrectly classified as either safe ground or definite hazards.

Temporal processing can reduce false detections and improve geometric consistency. As the AMR approaches a suspected pit or curb, new viewpoints provide additional observations of the terrain. Measurements from consecutive frames can be transformed into a common coordinate system and accumulated to determine whether the apparent discontinuity persists. A true terrain feature should generally become geometrically clearer, whereas transient noise or isolated missing measurements may disappear as additional evidence becomes available.

Vehicle motion must be considered during this temporal accumulation. Pitch and roll caused by acceleration, braking, suspension movement, or uneven terrain can change the apparent orientation of the ground relative to the sensors. IMU measurements and vehicle pose estimates can compensate for these effects so that a temporary chassis attitude change is not interpreted as a terrain discontinuity. Accurate sensor extrinsics and time synchronization are therefore particularly important for detecting small curbs or shallow depressions.

Negative obstacle severity should be evaluated relative to vehicle capability. Relevant parameters include wheel diameter, wheelbase, track width, ground clearance, approach and departure angles, suspension travel, center-of-gravity height, maximum allowable pitch and roll, and expected operating speed. The same detected depth or step height can represent a minor traversal cost, a region requiring reduced speed, or a completely non-traversable hazard depending on the physical configuration of the AMR.

The detected hazard can be inserted into an occupancy or traversability map using a conservative spatial representation. Definite pits or unsafe drop-offs may be marked as non-traversable, while uncertain regions can receive high traversal costs or remain unknown. Safety margins can expand the effective hazard boundary because localization error, perception uncertainty, and vehicle footprint make driving immediately beside a pit risky even when the geometric edge has been estimated accurately.

Detection distance is important for higher-speed outdoor platforms. A negative obstacle discovered only immediately before the front wheels reach it may leave insufficient distance for safe braking or avoidance. The perception system should therefore evaluate terrain sufficiently far ahead according to vehicle speed, stopping distance, sensing geometry, processing latency, and control response. Long-range detection may have greater uncertainty, requiring the system to refine the estimate continuously as the vehicle approaches.

Validation should include physically representative negative obstacles rather than relying only on ordinary obstacle-detection benchmarks. Test scenarios should contain pits of different dimensions, shallow and deep depressions, trenches, ascending and descending curbs, road edges, drainage channels, slopes, shadows, vegetation, and partially occluded terrain. Performance should be evaluated using detection range, boundary accuracy, false-positive rate, missed hazards, update latency, and the resulting ability of the AMR to stop or reroute safely.

For an outdoor AMR, negative obstacle detection is therefore a combined problem of geometry, visibility, uncertainty, vehicle embodiment, and temporal reasoning. Reliable operation requires the system to understand not only what the sensors detect but also where ground should have been observed and why it may be missing. When integrated with elevation mapping, terrain traversability, multi-sensor fusion, occupancy mapping, and local planning, negative obstacle detection enables safer navigation through complex off-road and semi-structured environments.

04.08. Dynamic Scene Update in Drivable Area Map [w/Code]

Dynamic scene updating maintains a drivable-area map as the environment changes around an autonomous vehicle or AMR. A region classified as free at one moment may become occupied by a pedestrian, vehicle, forklift, bicycle, or another robot, while a previously blocked region may later become available. The map must therefore represent the current operational state rather than preserve a static interpretation of past observations.

The update process begins with repeated acquisition of sensor observations from cameras, LiDAR, radar, and other perception sources. Each observation is associated with a timestamp and transformed into a common vehicle or map coordinate frame. Accurate calibration, localization, and time synchronization are essential because apparent environmental changes can otherwise be created by sensor misalignment, vehicle motion, or timing differences rather than by actual movement in the scene.

Ego-motion compensation separates movement of the sensing platform from movement of surrounding objects. As the vehicle translates and rotates, static structures appear to move relative to onboard sensors even though their world positions remain unchanged. Localization, odometry, and IMU information can transform consecutive observations into a consistent reference frame. After compensation, remaining spatial changes provide stronger evidence that objects or environmental conditions are genuinely dynamic.

Dynamic-object detection identifies observations that should not be permanently integrated into the static drivable-area representation. Vehicles, pedestrians, cyclists, forklifts, mobile equipment, and other robots can be detected through object detection, motion segmentation, optical flow, point-cloud comparison, or radar velocity measurements. Semantic information helps determine what type of object is moving, while geometric information determines its location, dimensions, and relationship to the currently available free space.

Object tracking provides temporal continuity between individual detections. Instead of treating every sensor frame independently, the system associates observations belonging to the same object over time and estimates position, velocity, heading, and potentially acceleration. Kalman filtering, probabilistic tracking, or learned tracking methods can support this process. Stable object identities allow the drivable-area map to distinguish persistent motion from isolated sensor noise and short-lived detection errors.

A dynamic drivable-area map should separate static and dynamic evidence whenever possible. Buildings, walls, fixed curbs, permanent barriers, and stable terrain can be retained with relatively long persistence, whereas observations associated with moving objects should be updated rapidly. Maintaining separate static and dynamic layers prevents temporary obstacles from contaminating the long-term representation while still allowing them to influence immediate local planning and collision avoidance.

Occupancy-grid updates can represent dynamic changes at the cell level. When an object enters a previously free cell, new measurements increase its occupancy probability. When the object leaves, subsequent free-space observations reduce that probability and eventually restore the region to free space. This clearing process is essential because a map that only adds obstacles will gradually accumulate stale occupancy and incorrectly reduce the available drivable area.

Temporal decay provides another mechanism for removing outdated information. Each observation can carry an age or timestamp, and its influence can decrease when it is not confirmed by newer measurements. Dynamic objects may use short decay periods, while static structures use much longer persistence. The decay rate should be selected carefully because aggressive clearing can erase temporarily occluded obstacles, whereas slow clearing can preserve objects that have already left the scene.

Occlusion complicates dynamic updates because a tracked object may temporarily disappear behind another vehicle, building, vegetation, or environmental structure. Immediate deletion of the object would incorrectly create free space. Tracking systems can maintain a predicted state for a limited period while representing increasing uncertainty. Visibility reasoning can determine whether the object should have been observable, helping distinguish true disappearance from temporary sensor occlusion.

Prediction extends dynamic mapping from the present state into the near future. Estimated velocity and motion history can be used to predict where pedestrians, vehicles, or other mobile agents may move during the planning horizon. Instead of representing only current occupied cells, the system can generate predicted occupancy or dynamic risk regions. This allows the planner to avoid trajectories that are currently free but are likely to become occupied before the autonomous platform reaches them.

Uncertainty should accompany both current and predicted dynamic states. Sensor noise, partial observations, ambiguous object associations, occlusion, and uncertain future behavior can all reduce confidence. A dynamic map can therefore store occupancy probability, object confidence, predicted motion uncertainty, and observation age. The planner can use these values to apply larger safety margins around uncertain moving agents rather than assuming that every estimated trajectory is exact.

Multi-sensor fusion improves dynamic-scene updating because different sensing modalities provide complementary temporal information. Cameras provide semantic classification and visual motion, LiDAR provides precise spatial geometry, and radar can directly measure relative radial velocity. Combining these observations helps distinguish moving objects from static structures and improves tracking when one sensor becomes unreliable because of lighting, weather, range, or partial occlusion.

The drivable-area map must also account for environmental changes that are not conventional moving objects. Doors may open or close, temporary barriers may be installed, equipment may be relocated, construction zones may change, and parked vehicles may appear or disappear. These changes occur more slowly than pedestrian or vehicle motion but still invalidate static assumptions. Map updating therefore requires multiple temporal scales rather than a single persistence rule for every environmental feature.

Dynamic updates must remain computationally efficient because perception, mapping, prediction, and planning operate continuously. Local rolling grids, region-of-interest updates, sparse data structures, parallel processing, and selective tracking can reduce computational load. Update frequency should reflect vehicle speed and stopping distance: faster platforms require more frequent and lower-latency map updates because stale information can become dangerous within a much shorter travel distance.

The updated representation is ultimately consumed by local planning and trajectory generation. Static obstacles define persistent geometric constraints, dynamic objects introduce time-dependent collision constraints, and uncertain regions contribute additional risk or cost. The planner can then select trajectories that remain within traversable space while accounting for where surrounding agents are expected to move. Dynamic scene updating therefore connects perception history directly with predictive motion planning.

Validation should evaluate not only detection accuracy but also the temporal correctness of the map. Important characteristics include object insertion latency, tracking consistency, stale-obstacle clearing time, occlusion handling, free-space recovery, prediction accuracy, and end-to-end planning response. Test scenarios should include crossing pedestrians, approaching vehicles, moving forklifts, stopped and restarting objects, temporary occlusions, disappearing obstacles, and changes to previously static environmental structures.

For outdoor AMRs, dynamic scene updating transforms the drivable-area map from a static geometric product into a continuously evolving world representation. By combining ego-motion compensation, object detection and tracking, occupancy updates, temporal decay, visibility reasoning, prediction, uncertainty, and multi-sensor fusion, the system maintains an actionable estimate of where the robot can move now and where it is likely to move safely in the near future.

04.09. Drivable Area Estimation Validation Metrics

Validation of drivable-area estimation determines whether the perception system produces spatial representations that are sufficiently accurate, stable, timely, and safe for autonomous navigation. Evaluation should extend beyond visual similarity between predicted and reference masks because the final output influences collision avoidance and trajectory planning. A useful validation framework therefore measures segmentation quality, geometric accuracy, uncertainty, temporal consistency, computational performance, and downstream navigation behavior.

Ground-truth construction is fundamental to meaningful evaluation. Reference drivable regions may be generated through manual annotation, high-precision LiDAR mapping, surveyed terrain, multi-sensor reconstruction, or combinations of these methods. The reference definition must clearly specify what constitutes drivable, non-drivable, and unknown space. For outdoor AMRs, annotations should also consider platform capability because a surface that is physically accessible to one vehicle may be unsuitable for another.

Intersection over Union, commonly abbreviated as IoU, is one of the most widely used metrics for evaluating drivable-area segmentation. It measures the overlap between the predicted drivable region and the ground-truth region relative to their union. A high IoU indicates strong spatial agreement, while a low value indicates missing or incorrectly added regions. Mean IoU can extend the measurement across multiple semantic classes such as drivable terrain, obstacle, vegetation, sidewalk, and unknown space.

Precision and recall provide complementary information that IoU alone cannot fully explain. Precision measures how much of the predicted drivable area is actually drivable, while recall measures how much of the true drivable area was successfully identified. Low precision can indicate dangerous false-free predictions in which unsafe regions are classified as traversable. Low recall indicates overly conservative perception that removes usable space and can unnecessarily restrict route planning or prevent mission completion.

False-positive and false-negative errors should be interpreted according to their safety consequences. Predicting an obstacle or unsafe terrain as drivable is generally more critical than incorrectly rejecting a safe region. Evaluation can therefore distinguish false-free and false-occupied errors rather than reporting only overall pixel accuracy. Safety-weighted metrics may assign larger penalties to errors near the vehicle, along the planned trajectory, or within the expected stopping region.

Boundary accuracy is important because navigation depends on the precise location of road edges, curbs, obstacles, pits, and terrain transitions. Two segmentation masks can achieve similar IoU values while producing very different boundary quality. Boundary distance, contour error, Hausdorff distance, or related geometric measurements can quantify how far estimated boundaries deviate from reference boundaries. Such metrics are particularly relevant when an AMR operates in narrow passages or close to physical hazards.

For BEV and occupancy-based representations, evaluation should be performed in physical coordinates rather than only in image pixels. Cell-level accuracy can measure agreement between predicted and reference free, occupied, and unknown states, while metric distance errors can quantify spatial displacement. Evaluation should report performance across distance bands because uncertainty generally increases farther from the sensors. Errors at five meters and fifty meters may have very different causes and operational consequences.

Traversability estimation requires metrics beyond semantic segmentation accuracy. Predicted terrain costs or traversable classes should be compared with actual vehicle capability and behavior. Validation can examine whether the robot successfully crosses a region, experiences excessive slip, violates slope or clearance constraints, or requires intervention. Cost prediction can also be compared with measured velocity, motor effort, vibration, or energy consumption to determine whether the traversability model reflects physical mobility.

Negative obstacle validation requires dedicated metrics because pits, ditches, drop-offs, and descending curbs differ from conventional positive obstacles. Detection rate, missed-hazard rate, boundary localization error, estimated depth error, and detection distance can be evaluated separately. Early detection is particularly important because a geometrically accurate pit estimate generated too late for the vehicle to stop provides little practical safety value.

Temporal consistency measures whether the estimated drivable area remains stable across consecutive observations when the environment itself is unchanged. Excessive frame-to-frame fluctuation can cause unstable planning even when individual frames have acceptable segmentation accuracy. Metrics can measure cell-state changes, boundary jitter, track persistence, or consistency after ego-motion compensation. Dynamic scenes should additionally evaluate how quickly newly occupied regions appear and how quickly cleared regions return to free space.

Dynamic-map validation should measure object insertion latency, stale-obstacle clearing time, tracking consistency, occlusion handling, and predicted occupancy accuracy. When a pedestrian or vehicle enters the drivable region, the map should represent the new hazard quickly enough for planning. When the object leaves, obsolete occupancy should disappear without excessive delay. These temporal characteristics directly affect how much usable free space the planner receives during real operation.

Uncertainty calibration should be validated independently from prediction accuracy. A well-calibrated system should report lower confidence in difficult conditions and higher confidence when observations are reliable. Reliability diagrams, calibration error, probabilistic scoring rules, or uncertainty-versus-error analysis can determine whether confidence values correspond to actual prediction quality. This is important because downstream planning may use uncertainty to enlarge safety margins or avoid poorly observed terrain.

Performance should be evaluated across environmental conditions rather than summarized only as one aggregate number. Day, night, rain, fog, shadows, wet pavement, vegetation, dust, gravel, slopes, construction areas, and sensor degradation can produce very different failure modes. Results can be divided by operational design domain conditions, terrain categories, distance ranges, and sensor availability so that weaknesses are visible instead of being hidden by average performance.

Computational metrics are also necessary because drivable-area estimation operates within a real-time autonomy pipeline. Processing latency, update frequency, memory consumption, accelerator utilization, and worst-case execution time affect whether perception results reach the planner before they become stale. Average frame rate alone is insufficient because occasional long delays may be safety-critical. End-to-end latency from sensor acquisition through map availability should therefore be measured under realistic computational load.

System-level validation evaluates whether perception quality actually supports successful autonomous behavior. Relevant outcomes include collision rate, safety intervention frequency, route completion, unnecessary stops, minimum obstacle clearance, path efficiency, and successful avoidance of non-traversable terrain. Closed-loop simulation and real-vehicle testing can reveal errors that offline segmentation metrics do not capture because perception, planning, control, vehicle dynamics, and environmental interaction influence one another.

A robust validation program should combine offline datasets, simulation, controlled proving-ground experiments, and representative field operation. Offline datasets enable repeatable algorithm comparison, simulation enables rare and hazardous scenarios, controlled testing provides measurable physical ground truth, and field trials expose the system to realistic complexity. Regression testing should preserve difficult historical cases so that software updates do not silently reintroduce previously corrected failures.

Drivable-area estimation validation should therefore be treated as a multi-dimensional engineering process rather than a single segmentation benchmark. IoU, precision, recall, boundary accuracy, occupancy consistency, traversability performance, uncertainty calibration, latency, and closed-loop navigation outcomes describe different aspects of system quality. Together, these metrics establish whether the estimated drivable area is not only visually correct but operationally useful, temporally reliable, and sufficiently safe for outdoor autonomous navigation.

04.10. Outdoor AMR Traversability Estimation Case Study

An outdoor AMR traversability estimation case study demonstrates how perception, vehicle modeling, mapping, and planning can be integrated to determine whether a mobile robot can safely cross complex terrain. Unlike conventional road navigation, the operational environment may include pavement, gravel, grass, soil, ramps, curbs, depressions, loose surfaces, and irregular ground. The objective is therefore not merely to detect free space, but to estimate whether the specific robot can physically traverse each region with acceptable risk.

The case-study platform can be considered a heavy outdoor AMR equipped with LiDAR, multiple cameras, IMU, GNSS-based localization, wheel odometry, and optional radar. Vehicle parameters such as wheel diameter, wheelbase, track width, ground clearance, suspension characteristics, mass, payload, and maximum allowable slope are incorporated into the perception model. These parameters convert generic terrain geometry into vehicle-specific mobility constraints rather than assuming that every geometrically open region is traversable.

The perception pipeline begins by synchronizing and calibrating sensor measurements before transforming them into a common vehicle coordinate frame. LiDAR provides three-dimensional terrain geometry, cameras contribute semantic surface information, and the IMU measures vehicle attitude and motion. GNSS and odometry provide localization and ego-motion information. Combining these observations creates a spatial representation in which terrain properties can be evaluated consistently as the AMR moves through the environment.

LiDAR measurements are first processed to estimate the ground surface and construct a local elevation map. Noise and isolated points are filtered before neighboring measurements are analyzed for height, slope, surface normals, and discontinuities. The resulting elevation representation identifies ramps, raised surfaces, depressions, curbs, and rough terrain. Unknown regions are retained when observations are insufficient rather than automatically being interpreted as safe free space.

Camera-based perception adds semantic information to the geometric representation. Image segmentation can distinguish pavement, gravel, grass, vegetation, soil, mud, water, and other visually identifiable surfaces. These semantic labels provide prior information about expected traction and surface behavior. Camera information alone does not determine traversability, however, because visually similar surfaces can have different physical properties and lighting or weather can significantly alter their appearance.

Geometric terrain features are then extracted from the local map. Longitudinal and lateral slope describe the orientation of the surface, while roughness represents smaller-scale height variation. Step height and discontinuity measurements identify curbs, rocks, ledges, potholes, and potential drop-offs. Each feature is evaluated over an area related to the vehicle footprint because terrain conditions underneath multiple wheels can produce roll, pitch, clearance problems, or unstable load distribution.

Vehicle constraints transform these terrain measurements into physical feasibility estimates. A slope is compared with climbing, braking, and lateral stability limits, while detected steps are compared with wheel radius, suspension travel, approach angle, departure angle, and ground clearance. Terrain beneath the complete footprint can also be evaluated for expected chassis roll and pitch. A region is rejected when one or more hard physical constraints exceed the allowable operating envelope of the platform.

Traversable regions that satisfy hard constraints are assigned continuous costs rather than being treated as equally desirable. Flat pavement may receive a low cost, while gravel, moderate slopes, rough surfaces, or uncertain terrain receive progressively higher costs. Expected wheel slip, vibration, energy consumption, and stability risk can contribute to this cost. The resulting traversability map therefore expresses both whether movement is possible and how difficult or risky that movement is expected to be.

Negative obstacles require a dedicated branch of the case-study pipeline. Missing ground observations, sudden downward elevation changes, unusual LiDAR ray patterns, and camera-based terrain boundaries are analyzed to detect pits, ditches, drainage channels, and drop-offs. Because missing measurements can also result from occlusion or poor reflectivity, suspected hazards are associated with uncertainty. Confirmed drop-offs are marked non-traversable, while ambiguous regions receive conservative costs until additional observations become available.

The AMR can improve its terrain estimate through proprioceptive feedback while driving. Wheel encoder velocity is compared with vehicle motion to estimate slip, while IMU acceleration, pitch, and roll indicate physical response to the terrain. Motor current or torque can provide evidence of increased resistance. If a surface predicted to be easy produces significant slip or vibration, its traversal cost can be increased, allowing the system to adapt perception-based expectations using actual vehicle-terrain interaction.

A representative mission can begin on paved ground and continue through gravel, grass, a moderate ramp, and an uneven off-road section. The perception system initially identifies several candidate routes toward the goal. A short route across rough terrain may have higher predicted cost, while a longer route over firm ground may offer greater stability. The planner uses the traversability map together with obstacle occupancy and mission constraints to compare these alternatives without relying solely on geometric path length.

As the vehicle proceeds, the map is continuously updated. A newly detected rock or parked object changes occupancy, while improved viewpoints refine slope and terrain-boundary estimates. Regions previously marked unknown may become traversable after additional observations, and temporary obstacles can be removed after they disappear. This continuous update prevents the planner from treating the initial environmental interpretation as permanent and allows navigation decisions to reflect the current scene.

Dynamic objects are maintained separately from persistent terrain properties. Pedestrians, vehicles, other robots, and moving equipment create time-dependent occupancy that influences immediate planning but should not permanently modify terrain classification. Object tracking and predicted motion provide short-term collision constraints, while the traversability layer describes longer-lasting physical properties of the ground. The planner combines both layers when selecting a locally safe trajectory.

Consider a situation in which the direct route contains a shallow ditch followed by loose gravel. The ditch depth may remain within wheel and suspension capability, but the combination of downward step, reduced traction, and lateral slope can produce an unacceptable risk. A purely geometric free-space detector might select the route because no conventional obstacle blocks it. The traversability system instead evaluates the combined physical conditions and can assign sufficient cost to favor a safer alternative.

Validation compares predicted terrain feasibility with actual vehicle behavior. Test runs record whether each region is successfully traversed together with wheel slip, vehicle attitude, speed reduction, motor effort, clearance, intervention, and route completion. Predicted traversability classes and costs can then be compared with these measurements. Incorrectly classifying dangerous terrain as traversable is treated as particularly important because such errors can directly expose the platform to immobilization, collision, or stability hazards.

Closed-loop testing evaluates the complete interaction between perception, mapping, planning, control, and vehicle dynamics. The AMR repeatedly encounters different surface types, slopes, curbs, roughness levels, negative obstacles, and uncertain regions while executing realistic missions. Successful validation requires more than accurate terrain classification: the vehicle must select feasible paths, maintain adequate clearance and stability, react to changing observations, and reach its destination without unsafe intervention.

This case study illustrates that outdoor AMR traversability estimation is fundamentally an embodied perception problem. Terrain cannot be classified independently of the robot that must cross it. By integrating multi-sensor perception, elevation mapping, semantic understanding, vehicle constraints, proprioceptive feedback, uncertainty, dynamic mapping, and closed-loop planning, the system transforms raw environmental observations into a continuously updated estimate of where the specific AMR can move safely and effectively.


Author: Richard Minsoo Go

LinkedIn: Richard Minsoo Go

Top comments (0)