DGT-Map: Directional Global Traversability Mapping Utilizing Multi-Task Learning for Heterogeneous Vehicles

I. INTRODUCTION

• Most existing maps assign one scalar cost per location, so planners cannot distinguish approach angles and may produce overly conservative or suboptimal paths.A vehicle may descend a slope safely but be unable to ascend it.Source: §I · p. 1
• Traversability also depends on vehicle geometry and dynamics, while many models are trained for one vehicle and may require retraining or fine-tuning for another.A short-wheelbase, high-clearance vehicle can cross a ridge that a longer vehicle collides with.Source: §I · p. 1, Figure 1 · p. 1
• DGT-MAP predicts global directional traversability costmaps from RGB-D observations using execution feedback and a shared terrain encoder with vehicle-specific heads.Joint training shares terrain experience across vehicles while retaining platform-dependent responses, which is useful when vehicle-specific data is limited.Source: §I · pp. 1–2
• DGT-MAP achieves the highest or tied-highest navigation success rate across challenging tasks and vehicle platforms versus geometric, binary, and direction-agnostic baselines.The evaluation integrates its directional maps into a Hybrid A* navigation stack and measures navigation success.Source: §I · p. 2
• The work contributes a direction-aware, vehicle-conditioned mapping framework, self-supervised multi-task learning, and closed-loop simulation evaluation of their effects on navigation success.The multi-task formulation is intended to improve predictions for vehicles with limited data.Source: §I · p. 2

A. Traversability Representation

• Binary labels and small category sets integrate with occupancy-grid planners but collapse terrain variability, while scalar cost fields provide smooth planner gradients yet assume isotropy.Some vehicle-frame approaches still assign one cost per spatial location and therefore do not fully represent direction-dependent traversability.Source: A. Traversability Representation · p. 2
• Directional traversability models condition cost on heading, preserving terrain asymmetries such as ascent versus descent for planning.Earlier directional work demonstrated geometric path-quality benefits but did not evaluate a planning and control stack's navigation performance.Source: A. Traversability Representation · p. 2

B. Traversability Estimation

• Geometric estimators use descriptors such as slope or roughness and can capture directional effects, but struggle to distinguish visually similar terrain with different physical properties and require hand-designed cost mappings.Self-supervised estimators instead pair terrain observations with proprioceptive outcomes, grounding labels in execution and naturally capturing direction-specific traversal.Source: B. Traversability Estimation · p. 2
• Most self-supervised traversability methods train independently for one vehicle, so deployment on another with different geometry or dynamics often requires retraining or fine-tuning.Multi-vehicle self-supervised training remains largely unexplored.Source: B. Traversability Estimation · p. 2

III. PROBLEM STATEMENT

• The target is a global traversability map whose terrain-induced cost depends on spatial location, travel direction, and vehicle platform.The map is evaluated by integrating it with a navigation system and measuring task success in reaching a goal.Source: §III · pp. 2–3

IV. METHODOLOGY

• DGT-MAP uses commanded-versus-actual motion discrepancy as self-supervision, then trains a shared terrain backbone with vehicle-specific heads before producing maps for a planner.The planner uses the predicted map to generate safe paths.Source: §IV · p. 3

A. Data Collection

• Data collection aggregates runs from four geometrically and dynamically distinct vehicles to broaden terrain coverage where less capable vehicles may not explore.The JEEP traverses the widest terrain range, including regions that may be inaccessible to less capable vehicles.Source: A. Data Collection · p. 3

1) Global Map:

• RGB-D observations from all runs are aggregated into a 0.4 m spatial grid, yielding a complete bird’s-eye-view representation independent of any single vehicle trajectory.The six-channel BEV representation includes RGB color and height statistics.Source: 1) Global Map · p. 3

2) Patches:

• Terrain patches are extracted in the vehicle’s local frame with forward aligned to the patch top, enabling direction-dependent traversability maps.Each patch measures 8.8 m × 2.8 m to cover vehicle footprints and surrounding terrain context.Source: 2) Patches · p. 3

3) Locomotion:

• Locomotion measures how closely realized motion matches commanded motion over a terrain segment, providing a continuous supervision signal for traversal difficulty.Higher L means easier traversal, opposite to the later traversability-cost convention.Source: 3) Locomotion · p. 3

4) Self-Supervised Label Generation:

• Sliding-window locomotion scores are paired with terrain patches across vehicles, naturally encoding terrain-dependent and vehicle-dependent variation.Different vehicles can traverse the same terrain and yield different outcomes.Source: 4) Self-Supervised Label Generation · p. 3

B. Training

• Because vehicles contribute imbalanced numbers of patches, the model jointly trains across vehicles while retaining vehicle-specific outputs rather than fitting separate models.The dataset imbalance arises from differences in traversal range.Source: B. Training · p. 3

1) Multi-Task Learning:

• A shared backbone learns terrain representations across vehicles while each vehicle-specific head captures its own locomotion response.For each training patch, only the head corresponding to the collecting vehicle contributes to the loss; other outputs are masked.Source: 1) Multi-Task Learning · pp. 3–4

2) Network Architecture:

• Two ResNet-18 branches process RGB and height-statistic channels separately before concatenation and vehicle-specific regression.The RGB branch uses ImageNet initialization; the geometric branch is trained from scratch.Source: 2) Network Architecture · p. 4

C. Costmap Inference

• Because planning needs dense costs for every location and direction, the inference pipeline turns single forward-patch predictions into global costmaps.The global BEV map is rotated to align each queried heading with the network’s forward axis.Source: C. Costmap Inference · p. 4, Figure 2 · p. 4

1) Spatial and Angular Discretization:

• Orientation is discretized into K heading bins, yielding a representation over spatial cells, heading, and vehicle.The experiments use K = 8.Source: 1) Spatial and Angular Discretization · p. 4

2) Patch Evaluation and Costmap Construction:

• Overlapping sliding-window predictions are averaged into a dense costmap, and rotating the map generates costs for other headings.The initial prediction layer represents forward travel; repeating the process across headings yields the directional map.Source: 2) Patch Evaluation and Costmap… · p. 4

D. Traversability-Aware Path Planning

• The planner keeps hard-collision checking separate from soft traversability preferences by using occupancy and directional-cost map layers.High directional cost can mark a difficult but feasible slope, whereas a cliff remains impassable.Source: D. Traversability-Aware Path Planning · pp. 4–5

1) Occupancy Map:

• The occupancy map marks cells occupied when their height gradient exceeds the slope threshold, when they are unobserved, and otherwise marks them free.The threshold is chosen not to eliminate terrain traversable in at least one direction; unobserved cells count as occupied during collision checks.Source: 1) Occupancy Map · p. 5

2) Directional Cost Lookup:

• The planner selects the nearest discretized heading bin for each successor’s travel direction and queries that directional cost layer.A nonlinear amplification penalizes high-cost terrain more strongly; β controls its strength and trades off path length against traversability.Source: 2) Directional Cost Lookup · p. 5

3) Hybrid A*:

• Each Hybrid A* expansion checks discretized motion primitives against occupancy and evaluates terrain cost at the successor cell.Occupied intermediate cells reject the primitive; each accepted edge incurs a distance cost plus terrain difficulty.Source: 3) Hybrid A* · p. 5, Figure 3 · p. 5

4) Controller Integration:

• Hybrid A* produces a global reference path that Nav2 MPPI tracks, with velocity commands converted to CARLA steering and longitudinal controls.Steering conversion uses the vehicle wheelbase; a PID controller regulates longitudinal motion.Source: 4) Controller Integration · p. 5

V. EXPERIMENTS

• The hill task contrasts a long gentle route with a shorter steep route that CARLACOLA can descend safely but cannot ascend.Trials provide only start and goal locations, requiring the planner to find a complete path without route constraints.Source: §V · p. 6
• DGT-MAP assigns the ridge low cost for the JEEP but substantially higher cost for the MERCEDESVAN, consistent with their different traversal capabilities.The ridge is traversable by the JEEP but immobilizes the MERCEDESVAN; a smoothed section offers the van a feasible crossing.Source: §V · p. 6, Figure 4 · p. 6
• The evaluation compares multi-task and single-task models on random start-and-goal navigation for JEEP and VOLKSWAGENVAN, with the van serving as a lower-data case.Trials use a terrain map similar to training and measure task success across multiple runs.Source: §V · p. 6

• DGT-MAP(ISOTROPIC) averages its eight predicted directional maps into one isotropic costmap.Source: A. Baselines · p. 6
• DGT-MAP(SINGLE-TASK) uses the DGT-MAP architecture but trains on data from only one vehicle.Source: A. Baselines · p. 6
• BINARY receives an occupancy map rather than a traversability costmap.Source: A. Baselines · p. 6
• SLOPE-BASED models locomotion from forward slope, fits its parameters to each vehicle’s training data, and converts predicted locomotion to terrain cost as 1 − Lsb.Its assumed locomotion peaks at θ0 and decreases toward zero at θmin and θmax; uphill and downhill decay rates may differ.Source: A. Baselines · pp. 6–7
• Table I covers two test maps and Table II random start-and-goal tasks; both define success as reaching within 5 m without immobilization for more than 10 seconds.The captions identify n as trials and k as successful runs.Source: Table I · p. 7, Table II · p. 7

Experiment 1

• The slope experiment highlights the value of preserving directional information when traversability varies with travel direction.Source: B. Results · p. 7
• DGT-MAP and SLOPE-BASED route around an infeasible uphill slope, while all methods take the same steep section downhill.This contrast shows DGT-MAP distinguishing feasible descent from infeasible ascent on the same terrain.Source: Figure 5 · p. 7, B. Results · p. 7

Experiment 2

• On the ridge map, DGT-MAP routes through the smoothed crossing, while SLOPE-BASED crosses the sharp bump that it does not account for.DGT-MAP and its isotropic variant achieve the highest success rates in this scenario; explicit heading conditioning provides less benefit than on the hill.Source: Figure 5 · p. 7, B. Results · p. 7

Vehicle-specific costmaps

• The JEEP assigns the sharp bump low cost, while the MERCEDESVAN assigns it substantially higher cost, matching their differing traversal capabilities.Source: B. Results · p. 7

Experiment 3

• Success rose from 64.0% to 70.0% for the JEEP and from 35.4% to 54.5% for the VOLKSWAGENVAN with multi-task versus single-task training.Source: B. Results · p. 7

VI. CONCLUSION

• DGT-MAP combines heading-aligned terrain representations, locomotion-derived supervision, and multi-task learning to capture directional and vehicle-specific traversability.Source: §VI · p. 7
• DGT-MAP improves navigation performance over evaluated binary, slope-based, isotropic, and single-task baselines when integrated with Hybrid A* and closed-loop control.Source: §VI · p. 7
• Experiments show that multi-task training improves performance over single-task training for both evaluated vehicles, while representing traversability by vehicle and direction helps planners distinguish where terrain is difficult.Source: §VI · p. 8

VII. LIMITATIONS & FUTURE WORK

• Training and evaluation are entirely in simulation, whose physics, terrain models, and visual diversity do not fully capture real off-road conditions.Source: §VII · p. 8
• Vehicle-specific output heads assume the target vehicle was represented during training, and zero-shot generalization to unseen vehicles was not evaluated.Source: §VII · p. 8
• Experiments do not isolate RGB's contribution, and the ridge result establishes performance beyond the slope-based baseline but not geometry as a whole.Source: §VII · p. 8
• Maps use simulator ground-truth localization; real deployment would require a localization framework whose noisier estimates could affect performance.Source: §VII · p. 8
• Directional costmaps are generated offline before planning, inference time grows linearly with heading-bin count, and the costmap does not inform low-level control.Source: §VII · p. 8