Blog

Traversability: How a Robot Decides Whether Terrain Is Drivable

Tom Cosgrove,

Reading Room: The Fictionlab Blog

TL;DR: Traversability is a robot’s estimate of whether a patch of terrain can be safely crossed given its geometry, surface properties, and the robot’s own drivetrain limits. In a typical ROS 2 stack, traversability is computed from elevation and semantic data, converted into a cost layer, and passed to a planner that either avoids untraversable cells or routes through lower-cost ones. Good traversability estimation combines geometric analysis (slope, roughness, step height) with semantic analysis (grass vs. mud vs. gravel) and matches those values against the platform’s measured capabilities.

Why Traversability Matters for a Mobile Robot on Real Terrain

A binary obstacle map can work well indoors. Outdoors, it breaks down. Tall grass can return LiDAR points yet still be drivable. A 12 cm curb can stop a small UGV and may be difficult to detect reliably at some sensor resolutions or viewing angles. A wet clay slope may look similar to dry gravel in RGB imagery, yet one may be drivable and the other is not. This is the core problem traversability estimation solves: deciding, per terrain patch, whether the robot can cross it, how costly the crossing is, and what risk it carries.

For field robots like the Leo Rover operating in agriculture, research plots, or construction sites, this decision runs continuously as the robot moves. Get it wrong toward the optimistic side and the robot tips, high-centers, or gets stuck. Get it wrong toward the pessimistic side and the robot refuses valid paths, wasting energy and failing missions.

What Traversability Actually Means in Practice

Traversability is often expressed as a scalar value per cell in a 2.5D grid, for example normalized between 0 (fully drivable) and 1 (impassable), sometimes with an “unknown” state. The exact scale and direction of the value depend on the system. That value is derived from several inputs:

  • Slope – the local inclination of the terrain surface, computed from the gradient of an elevation map.
  • Roughness – the variation in elevation within a window or residual error from a local plane fit, capturing bumpiness the suspension must absorb.
  • Step height – the maximum vertical discontinuity between nearby cells, capturing curbs, rocks, and roots.
  • Surface type – a semantic class such as asphalt, grass, sand, mud, or water, typically from a segmentation network.
  • Drivetrain properties – the platform’s max climb angle, ground clearance, wheel diameter, and available torque, which help set the thresholds against which the above are compared.

The Leo Rover, for reference, has 4WD with 130 mm wheels, about 80 mm of ground clearance, and a documented ability to climb slopes up to about 45 percent (approximately 24 degrees) on firm surfaces. These values provide an initial basis for thresholds in a traversability function, but the thresholds should be validated in representative conditions. A different platform demands different thresholds.

How Geometric Traversability Is Computed

Geometric methods rely only on 3D structure, usually from a depth camera, stereo pair, or LiDAR. Common ROS 2 building blocks include grid_map and the elevation_mapping_cupy package from ETH Zurich, which fuses point clouds into a probabilistic 2.5D elevation grid on GPU.

From that elevation grid, a traversability layer is typically built with these steps:

  1. Estimate a local surface normal per cell by fitting a plane to neighboring cells.
  2. Derive slope as the angle between that normal and the vertical axis.
  3. Compute roughness as residual variance from the fitted plane.
  4. Compute step height as a maximum elevation difference within a local neighborhood, for example among 8-connected cells.
  5. Combine these into a weighted score, then apply thresholds derived from the robot’s physical limits.

A simple weighted form looks like T = w1 * f(slope) + w2 * f(roughness) + w3 * f(step), where each f is a sigmoid or piecewise function that returns near 0 below the platform’s limit and saturates near 1 above it. Weights are usually tuned empirically on the target platform.

Geometric methods are fast, deterministic, and require no training data. Their weakness is that they cannot directly determine whether a soft surface will provide sufficient traction or support.

How Semantic Traversability Fills the Gap

Semantic traversability uses camera images (sometimes fused with LiDAR intensity) and a neural network to classify each pixel or point into a terrain class. Datasets like RELLIS-3D and RUGD are widely used for training off-road segmentation models. Each class can then be mapped to an expected friction range, sinkage risk, or directly to a traversability cost.

The difference between geometric and semantic assessment is straightforward. Geometric analysis answers “can the shape be climbed?” Semantic analysis provides a prior about whether the surface is likely to support the wheels. Neither is sufficient alone. A grass field is geometrically flat but semantically variable. A rocky slope is geometrically hostile regardless of material.

Modern stacks fuse both. Recent work such as Wild Visual Navigation from ETH Zurich learns traversability online from the robot’s own experience, associating visual features with proprioceptive feedback. That approach can adapt to novel environments without requiring pre-labeled terrain data for every deployment environment.

How Traversability Reaches the Costmap and the Planner

In a Nav2 pipeline, traversability can be injected as a custom costmap layer. The flow looks like this:

  1. Sensors publish point clouds and images.
  2. An elevation mapping node builds a 2.5D grid_map.
  3. A traversability node reads the grid_map, applies geometric and semantic analysis, and publishes a cost grid.
  4. A Nav2 costmap plugin subscribes to that grid and writes values into the global or local costmap.
  5. The planner (such as SmacPlanner or NavFn) and the local controller (such as MPPI) use those costs when configured to do so.

Cells at or above a configured lethal threshold behave like obstacles. Cells in the middle range bias the planner toward smoother, safer routes without forbidding rougher ones. This is why traversability is more powerful than binary occupancy: it lets the robot prefer a longer smooth path over a shorter bumpy one, using a single cost function the planner already understands.

The Decision Process, Step by Step

The end-to-end decision a robot makes per control cycle can be summarized as a chain:

  1. Perceive: LiDAR and camera produce point clouds and images at sensor-dependent rates.
  2. Map: Points are registered into an elevation grid; images are segmented into terrain classes.
  3. Score: Each cell receives geometric metrics (slope, roughness, step) and a semantic class.
  4. Fuse: A traversability function combines these into a cost, using limits validated for the robot and operating conditions.
  5. Plan: The global planner searches for a low-cost path; the local controller checks feasibility over a short horizon.
  6. Verify: IMU, wheel odometry, and motor current can confirm whether the surface behaves as predicted; deviations can update the map or model.

The last step matters. A robot that never checks whether its predictions matched reality will repeat the same mistakes. Feedback from proprioception is what turns a static traversability estimator into an adaptive one.

When Geometric Analysis Is Enough and When You Need Semantics

For structured outdoor environments (paved paths, warehouse yards, well-maintained gravel roads), geometric traversability from a depth sensor is often sufficient. The surface properties are uniform enough that shape dominates the decision.

For unstructured field work, such as agricultural rows, forest floors, or research sites tracked in research applications, semantic information is often valuable. Vegetation, mud, standing water, and loose soil can look benign geometrically but change drivability drastically. The same applies to agriculture deployments where crop rows, irrigation channels, and seasonal soil moisture shift the traversable set week to week.

Common Failure Modes to Watch For

Even a well-tuned traversability stack has predictable weak spots:

  • Negative obstacles: holes and drop-offs are hard to see with forward-facing sensors until the robot is close.
  • Tall vegetation: grass and thin branches can register as obstacles in raw point clouds even when they are traversable, although dense vegetation can still impede the robot.
  • Reflective, transparent, or low-return surfaces: water and some dark materials can degrade the measurements of active depth sensors.
  • Domain shift: semantic networks trained on one biome can fail in another.
  • Sensor mounting: a LiDAR mounted too low creates blind zones near the wheels where step obstacles hide.

Addressing these usually requires sensor redundancy, conservative fallback behaviors, and periodic retraining or online adaptation of the semantic component.

Frequently Asked Questions

What is the difference between traversability and occupancy?

Occupancy represents whether a cell is believed to contain an obstacle, often as a probability that can be thresholded into free or occupied. Traversability represents how difficult or risky it is to cross a cell, often as a continuous cost. Occupancy handles walls and static obstacles well. Traversability handles slopes, rough ground, and variable surfaces.

Which ROS 2 packages are used for traversability estimation?

Common building blocks include grid_map for 2.5D mapping, elevation_mapping_cupy for GPU-accelerated elevation fusion, and custom Nav2 costmap plugins that convert traversability grids into planner-ready costs. Research packages like Wild Visual Navigation add learned components.

Do you need LiDAR, or is a depth camera enough?

A depth camera such as an Intel RealSense or a stereo pair is often sufficient for small UGVs at low speeds and short planning horizons. LiDAR generally extends usable range and can be more robust in low-texture scenes; performance in direct sunlight depends on the specific sensor technology. Many outdoor stacks fuse both.

How does a robot handle surfaces it has never seen?

Two approaches: conservative defaults (unknown surfaces get high cost until proven otherwise) or online learning (the robot drives cautiously and updates its model from proprioceptive feedback). Wild Visual Navigation is one example of the second approach.

Can traversability be computed without machine learning?

Yes. Purely geometric methods using slope, roughness, and step height require no training data and can run in real time on modest hardware. They are a standard starting point and are often enough for structured environments.

How is traversability validated in the field?

Typical validation compares predicted traversability against measured outcomes: IMU vibration levels, wheel slip, motor current, and success rate on labeled test transects. Mismatches drive threshold retuning or model updates.

What slope can a small UGV realistically climb?

It depends on drivetrain, weight distribution, wheel material, and surface. The Leo Rover is documented to handle slopes up to about 45 percent on firm ground. Loose or wet surfaces reduce that limit substantially, which is why traversability functions must incorporate surface type, not just angle.

Build and Test Traversability on a Real Platform

Traversability estimation is only as good as the platform you validate it on. If you are developing outdoor autonomy stacks and need a ROS 2-compatible robot to prototype elevation mapping, costmap layers, and planner tuning against real terrain, explore the Leo Rover as a field-ready development base.


Read more