PaperlayerView exact revision on arXiv

Source-linked AI summary

Learning Ground Traversability from Simulations

R. Omar Chavez-Garcia, Jerome Guzzi, Luca M. Gambardella, Alessandro Giusti

arXiv:1709.05368v3cs.RO

TL;DR

Predicting where a ground robot can travel on unstructured terrain is difficult because traversability can depend on the robot and travel direction. The paper trains a CNN on simulated, procedurally generated heightmap patches to estimate oriented traversability. The authors report that the CNN outperforms baseline and feature-based estimators on synthetic and real-world heightmaps, while noting limits for terrain whose physical properties are not represented in training.

  • Problem

    On unstructured terrain, traversability can depend on travel direction and robot type, and terrain difficulties may be hard to anticipate.

    Source: §I · p. 1

  • Method

    The approach trains a CNN to classify oriented heightmap patches using labeled examples generated by simulating a robot on procedurally generated terrains.

    Source: §I · p. 1, 1) Heightmap Generation · p. 3

  • Results

    The CNN estimator outperforms baseline and feature-based approaches on synthetic and real-world heightmaps, though performance is lower on the real Sullens, Gravelpit, and Quarry datasets than on synthetic evaluation data.

    Source: A. Classification Results · p. 4

  • Takeaways & Limitations

    The resulting oriented traversability maps can support path planning that trades off path length against the probability of traversability.

    Source: §V · p. 7

    The method assumes terrain shape alone determines traversability and uses only simulated training data, limiting its suitability for sand, mud, wet or slippery terrain and leaving simulation-to-reality mismatch as a concern.

    Source: §V · p. 8

Abstract

from arXiv · show

Mobile ground robots operating on unstructured terrain must predict which areas of the environment they are able to pass in order to plan feasible paths. We address traversability estimation as a heightmap classification problem: we build a convolutional neural network that, given an image representing the heightmap of a terrain patch, predicts whether the robot will be able to traverse such patch from left to right. The classifier is trained for a specific robot model (wheeled, tracked, legged, snake-like) using simulation data on procedurally generated training terrains; the trained classifier can be applied to unseen large heightmaps to yield oriented traversability maps, and then plan traversable paths. We extensively evaluate the approach in simulation on six real-world elevation datasets, and run a real-robot validation in one indoor and one outdoor environment.

I. INTRODUCTION

Outdoor terrain traversability can depend on direction, robot, and terrain features, making it difficult to construct useful maps in advance. The paper presents a machine-learning framework that predicts oriented traversability from heightmap patches and applies the estimates to path planning.

I. INTRODUCTION

  • Outdoor traversability may depend on direction and robot type, while terrain features such as holes, power cords, and speed bumps can create difficulties that are hard to anticipate.For example, a wheeled robot may descend but not ascend a slope, and a tall robot may capsize when crossing it sideways.Source: §I · p. 1
  • A patch is locally traversable in a direction when the robot, placed at its center, can proceed that way for a short distance under its low-level controller.The patch is centered at the robot’s position, oriented with the robot facing right, and represented by pixel heights for binary CNN classification.Source: §I · p. 1
  • Training examples come from simulations on procedurally generated terrains, where successful and unsuccessful progress supplies labeled patches for applying the learned classifier densely to unseen heightmaps.Evaluating many points and orientations with the classifier is orders of magnitude faster than simulating the robot at each point.Source: §I · p. 1
  • The framework combines procedural generation of training heightmaps, a traversability classifier, and path-planning application, with validation on real-world heightmaps.The classifier is trained for a specific robot and is intended to apply across wheeled, tracked, legged, and snake-like locomotion.Source: §I · p. 1

II. RELATED WORK

Prior traversability work uses appearance or terrain geometry and obtains labels through robot experience, human annotation, or other strategies. This paper instead learns from raw heightmap geometry using a CNN trained on simulated examples from procedurally generated terrains.

Inputs

  • Related approaches use appearance cues directly or segment images into terrain classes, while geometry-based methods derive elevation maps and use simulation, handcrafted rules, or learned classifiers.Appearance-based approaches may use generic visual features with standard classifiers or CNNs on raw images.Source: §II · p. 2
  • The approach feeds raw elevation patches to a CNN, which learns robot-specific terrain patterns without prior feature assumptions and outperforms a feature-based approach in the reported experiments.The authors describe this as the first application of deep learning to heightmap data for traversability estimation.Source: §II · p. 2

Source of training data

  • The paper relies exclusively on simulated data from procedurally generated terrains, enabling cheap generation of large labeled datasets for data-hungry deep-learning models.The authors identify this as an untried strategy in traversability literature at the time of the paper.Source: §II · p. 2
  • Real-robot experiments in two environments support transfer from simulation on solid, non-slippery terrain, but the classifier may perform poorly on sandy, unstable, or slippery terrain unless those properties are simulated during training.The experiments refute the concern that procedural terrain shapes necessarily prevent good performance on real heightmaps.Source: §II · p. 2

III. TRAVERSABILITY ESTIMATION FRAMEWORK

The framework section introduces the proposed traversability-estimation pipeline and previews its simulation setup, dataset generation, and classification methods.

III. TRAVERSABILITY ESTIMATION FRAMEWORK

  • Figure 1 illustrates the framework; the section then covers simulation setup, training and evaluation dataset generation, and classification approaches.Source: §III · p. 2

A. Traversability from Simulation

The simulation framework generates training data by running a modeled robot on procedurally generated terrain. The generated heightmaps combine terrain patterns and added obstacle-like features to produce varied training conditions.

A. Traversability from Simulation

  • The simulated Pioneer 3-AT traverses heightmap-defined uneven terrain at a constant 15 cm/s, and its recorded trajectory supplies traversability information.The setup uses Gazebo with the ODE physics engine.Source: A. Traversability from Simulation · pp. 2–3

1) Heightmap Generation:

  • The authors choose procedural generation to create arbitrarily large training datasets and test whether classifiers trained on synthetic heightmaps work on real heightmaps.Source: 1) Heightmap Generation · p. 3
  • Thirty 512 × 512 training terrains represent 10 m × 10 m areas and combine simplex-noise components with varied periods, heights, and spatially modulated weights.The component combinations span flat ground, rocks, smooth hills, and rocky hills.Source: 1) Heightmap Generation · p. 3
  • Additional heightmap transformations create holes, steps, bumps, and rail-like indentations.Source: 1) Heightmap Generation · p. 3

2) Simulation Process:

Simulation trajectories provide labeled, robot-centered heightmap patches, while training and evaluation maps are generated or collected across varied terrains and real-world scenarios.

2) Simulation Process:

  • Random starts on 30 terrains send the robot forward without steering until it reaches a map edge or remains stuck, after which it is respawned.Recorded poses and patches are labeled traversable only when a future pose lies beyond threshold d and aligns with the robot’s orientation; otherwise they are non-traversable.Source: 2) Simulation Process · p. 3
  • With T = 1 s and d = 0.12 m, the robot traverses smooth trajectory patches until blocked by a nearly robot-height bump; the first patches are labeled traversable and the last non-traversable.Source: 2) Simulation Process · p. 3

B. Dataset Generation

  • The real-world evaluation covers six scenarios and 350k samples, including three public heightmaps, an ETH-ASL map, and qualitative Bars and Slope environments.All datasets use T = 1 s, d = 12 cm, and patches equivalent to 1.2 m × 1.2 m.Source: B. Dataset Generation · pp. 3–4
  • The training terrains include varied obstacles, while the Slope evaluation map is an irregular grass incline bordered by a smooth uphill walkway.The Slope map was manually mapped with a Tango RGB-D sensor; Fig. 2 shows four of the 30 procedural terrains.Source: B. Dataset Generation · p. 4, Figure 2 · p. 4

C. Training Traversability Classifiers

The paper compares feature-based classification with a CNN that classifies heightmap patches directly, using handcrafted terrain descriptors or learned features, respectively.

C. Training Traversability Classifiers

  • The classifier predicts whether a heightmap patch is traversable, using either descriptive features with standard statistical classifiers or a CNN operating on raw input.The comparison also includes a dummy classifier that always predicts the training set’s most frequent class.Source: C. Training Traversability Classifiers · p. 4
  • The feature-based approach classifies 324-dimensional HOG descriptors with a 10-tree Random Forest.HOG captures cues including local steepness and step height.Source: C. Training Traversability Classifiers · p. 4
  • The CNN learns problem-specific features from 60 px × 60 px heightmaps through convolutional, pooling, and fully connected layers.It is trained for 100 epochs using categorical cross-entropy and Adadelta; evaluated architecture variations had minimal performance differences.Source: C. Training Traversability Classifiers · p. 4

IV. EXPERIMENTAL RESULTS

The CNN outperforms the baseline and feature-based estimators on synthetic and real-world heightmaps, though performance is lower on several real datasets; a real-robot bar test also shows low predicted traversability for obstacles the robot cannot cross.

IV. EXPERIMENTAL RESULTS

  • The evaluation compares the CNN and feature-based classifiers with a dummy classifier that always predicts the most frequent training class.Performance is reported across the evaluation datasets in Table I, which is not included among the supplied passages.Source: §IV · p. 4

A. Classification Results

  • The CNN outperforms both the baseline and feature-based approaches on synthetic and real-world heightmaps, with lower performance on Sullens, Gravelpit, and Quarry.The authors suggest patterns such as narrow passages, tight rails, and low barriers may be underrepresented in procedural training terrains; many real-terrain features are nevertheless correctly classified.Source: A. Classification Results · p. 4
  • The real Pioneer 3-AT rarely or never crosses the two tested bars, and the classifier assigns both very low traversability probabilities.It estimates about 50% traversability when the robot can climb a bar with only one wheel.Source: A. Classification Results · pp. 4–5

B. Traversability Visualization on the Quarry Dataset

Quarry traversability estimates vary by travel direction: the estimator marks the main road traversable but can label steep slopes traversable downhill, while the minimum-over-directions view is compact but omits directional detail.

B. Traversability Visualization on the Quarry Dataset

  • The estimator correctly marks the approximately 2 m-wide main road traversable, but steep slopes may be non-traversable uphill and dangerously marked traversable downhill.The authors note the definition tests whether the robot can proceed, not whether the motion is safe; slopes are always classified traversable downhill and sometimes transversally.Source: B. Traversability Visualization on the… · p. 5
  • Streets, sidewalks, and garden entries in Sullens are estimated traversable, while challenging uphill slopes receive average traversability confidence.The quarry’s traversable areas include roads, flat terrain, short bumps, and slopes.Source: B. Traversability Visualization on the… · p. 5

C. Computational cost

Training requires terrain generation, robot simulation, and CNN training, while inference densely evaluates oriented patches across a map. The authors report that inference is nearly 1000 times faster than the corresponding simulation workload.

C. Computational cost

  • Generating terrains takes a few milliseconds, producing the 450k-sample dataset takes about one day, and CNN training takes about one hour on a Tesla K80 GPU.The simulation stage dominates this training pipeline's reported time.Source: C. Computational cost · pp. 5–6
  • 3200 patches per square meter take 350 ms to evaluate on an Intel i7 CPU, and inference can be parallelized on a GPU.The estimate uses a 10 cm stride and 32 orientations; inference time is independent of patch content.Source: C. Computational cost · p. 6
  • Simulating traversal of 1 m^2 takes about 320 seconds at 10 cm/s and 1× real-time simulation speed, almost 1000× the time required by the proposed approach.Simulation speed varies with terrain complexity, from 0.1× real time on very complex maps to 10× on flat areas; the setup was not optimized for speed.Source: C. Computational cost · p. 6

D. Path Planning on Probabilistic Traversability Maps

The paper plans paths using probabilistic traversability estimates, combining segment probabilities under stated assumptions and accounting for places where the robot cannot turn in place. It presents length–traversability trade-offs and real-robot trials, including a verified route and a predicted route that could not be teleoperated successfully.

D. Path Planning on Probabilistic Traversability Maps

  • Treating every non-traversable patch as causing a stuck, crashed, or unrecoverable robot is pessimistic because some blocked attempts allow turning back and taking another route.The paper notes a wall as an example of a patch a robot cannot pass but may retreat from.Source: D. Path Planning on Probabilistic… · p. 6
  • Path probability is modeled as the product of segment probabilities, assuming independence along the path and that the robot can rotate in place.Because robots may not turn in place on unstructured terrain, a turnability classifier estimates the probability of rotating at least 45° in either direction; graph edges use traversability or turnability estimates.Source: D. Path Planning on Probabilistic… · p. 6
  • Pareto-optimal routes trade off path length against traversability: the shortest route may have low probability, while the most traversable route may be unnecessarily long.The figure shows route colors from red (non-traversable) to green (surely traversable) and a length–traversability trade-off for the bottom source–target pair.Source: D. Path Planning on Probabilistic… · p. 7, Figure 7 · p. 7
  • The maximal-traversability route to the Slope map's square goal followed a smooth sidewalk ramp and was verified traversable, while the route to the star goal had probability 0.21 and was blocked by a bump.The star-goal route could not be successfully teleoperated; it avoided excessively steep or rugged areas but still encountered the bump.Source: D. Path Planning on Probabilistic… · p. 7, Figure 8 · p. 7
  • The reachability map assigns each target the maximum traversal probability over paths from the robot's current pose.The map visualizes areas reachable from the initial pose.Source: D. Path Planning on Probabilistic… · p. 7, Figure 8 · p. 7

V. DISCUSSION AND CONCLUSIONS

The framework uses simulation-trained heightmap classifiers to estimate oriented traversability maps on large, unseen real-world terrains and supports path planning. Its scope is limited by omitted dynamics, shape-only assumptions, and simulation-only training.

V. DISCUSSION AND CONCLUSIONS

  • Simulation-trained classifiers estimate oriented traversability maps efficiently and accurately on large, unseen real-world terrains across robot locomotion modalities.The framework casts traversability estimation as heightmap classification.Source: §V · p. 7
  • The maps support path planning that trades off minimizing path length against maximizing the probability of traversability.Source: §V · p. 7
  • The method omits robot dynamics, limiting its intended scope to slow robots on rugged terrain where dynamic effects have negligible impact.Approach speed is one example of an omitted dynamic factor.Source: §V · p. 7
  • Assuming terrain shape alone determines traversability is inadequate on sand, mud, wet, or slippery terrain.The shape-only approximation is described as acceptable in some settings, such as solid ground where friction is not key.Source: §V · p. 8
  • Training only on simulated data is a limitation because simulations rarely reproduce the real world accurately.Source: §V · p. 8
  • Future work proposes training with real data by associating traversed heightmap patches with success or failure outcomes.Those outcomes could accumulate experience for training new models or refining existing ones, including learning effects of soft ground or slipping.Source: §V · p. 8
  • Shape-related exceptions could be addressed with separate models for terrain types or one model that also uses terrain appearance.The passage gives paved and muddy slopes with the same shape as an example of why appearance may help distinguish terrain types.Source: §V · p. 8

Related papers

Loading 1709.05368v3…