Off-road traversability is direction-dependent and vehicle specific, yet most global maps assign a single isotropic cost to each location. Existing learned estimators are also commonly trained independently for each vehicle; this preserves vehicle-specific behavior but prevents vehicles from sharing common terrain representations. DGT-MAP addresses both limitations through a self-supervised framework that learns global, directional, and vehicle-conditioned traversability costmaps from RGB-D observations and locomotion signals. A shared multi-task backbone learns common terrain features across training vehicles while vehicle-specific prediction heads preserve platform-dependent responses. At inference, DGT-MAP produces a heading-indexed costmap that can be used by a direction-aware planner. We evaluate DGT-MAP in simulation by integrating it into a Hybrid A* navigation stack and measuring downstream task success on challenging terrains, including slopes that are traversable downhill but not uphill and a ridge obstacle that is traversable by some vehicles, but not by others. Across evaluated tasks, DGT-MAP achieves the highest or tied-highest navigation success rate when compared against geometric, binary, and learned direction-agnostic baselines.
Figures & tables
Fig. 1 : Top: The traversability of terrain depends on the driving direction of the vehicle. For example, a vehicle might be able to go down a hill (left) but not up (right) due to limited horsepower. Bottom: Similarly, the traversability depends on the type of vehicle. For example, vehicles with a long wheel-base, like the van shown on the right, can collide when driving over ridge terrain, while a vehicle with higher ground clearance and a shorter wheel-base can drive over it (left).
Fig. 2 : DGT-Map training and inference pipeline. Top: Multi-vehicle self-supervised training. RGB-D observations from multiple vehicles are fused into a global BEV map, from which heading-aligned terrain patches are extracted along the traversed path. Locomotion labels are computed from the discrepancy between desired and recorded motion. A shared CNN is trained to predict these labels from terrain patches, with vehicle-specific heads preserving platform-dependent traversability behavior. Joint training across vehicles enables shared terrain representation learning. Bottom: Directional global traversability map generation at inference. The global BEV map is rotated to align a queried heading with the network’s forward axis. A sliding window extracts patches, which are processed by the trained CNN, and overlapping predictions are aggregated to form a costmap for a single-direction. Repeating this procedure across headings yields the full directional traversability map used for planning.
Fig. 3 : Traversability-aware planning. (a) When expanding from grid cell (i,j) , the planner computes the travel direction to neighboring cells (in′,jn′) for n=1,2,3,4 and queries the corresponding directional cost layer. (b) Each candidate motion primitive is discretized and checked against the occupancy map to ensure collision-free feasibility. (c) Four of the K=8 directional costmap layers ( θ=0∘,90∘,180∘,270∘ ) illustrating how traversability varies with heading; differences across layers capture anisotropic terrain effects such as ascending versus descending slopes.
Fig. 4 : Top Left: Hill Map showing Route A, a roundabout route with a gentle slope, and Route B, a direct route containing a steep section. Top Right: Ridge Map showing Route A, a path that avoids the sharp ridge, and Route B, a path that crosses the ridge. Route B is not traversable by some vehicles. Bottom Left: Traversability map of the Jeep , note how the ridge is shown as traversable. Bottom Right: Traversability map for the MercedesVan , note how the ridge has a low traversability.
Hill ( Cola ) ( n=20 )
Ridge ( MerVan ) ( n=50 )
Method
k
SR (%)
k
SR (%)
dgt-map
16
80
47
94
dgt-map(Iso)
10
50
45
90
dgt-map(ST)
5
25
14
28
Slope-Based
16
80
26
52
Binary
10
50
20
40
TABLE I : Navigation Success Rate on two test maps. n is the number of trials; k is the number of successful runs. The best result in each environment is bolded . Success is defined as reaching within 5 m of the goal without getting immobilized for greater than 10 seconds.
Jeep ( n=100 )
VwVan ( n=99 )
Method
k
SR (%)
k
SR (%)
dgt-map
70
70.0
54
54.5
dgt-map(Single-Task)
64
64.0
35
35.4
TABLE II : Navigation Success Rate on Random Start-and-Goal Tasks. DGT-Map (multi-task) is compared against its single-task variant on a map that is similar to the training environment but was not seen during training. n denotes the number of trials; k the number of successes. The best result per vehicle is bolded . Success is defined as reaching within 5 m of the goal without getting immobilized for greater than 10 seconds.
Fig. 5 : Planned paths in CARLA using 5 different costmaps with Hybrid A*. Right: Path planned for the CarlaCola to ascend from base to summit. dgt-map (blue) and Slope-Based (red) route around the steep slope via the longer path, correctly identifying that ascent through the steep slope is infeasible. The remaining methods attempt (and fail) to climb the steep slope directly. Top left: Paths planned to descend from the summit to the base. All methods, including dgt-map , route through the steep section, confirming that dgt-map distinguishes feasible descent from infeasible ascent on the same terrain. Bottom left: Ridge map scenario. dgt-map routes through the smoothed region, while the other methods end up taking the less traversable path through the sharp bump.
Autonomous navigation across large off-road environments remains a challenging problem. Onboard sensors perceive only the immediate surroundings, yet safe and efficient routes depend on terrain features that extend well beyond the sensor horizon. Geo-spatial data sources such as satellite imagery, aerial LiDAR, and vector maps can close this gap, but learning traversability from them is difficult: dense labels are unavailable at scale, and existing methods rely on short-range sensing. We propose an efficient formulation that learns a continuous traversability map from overhead data, supervised directly by human-driven GPS trajectories and shaped by supervised geometric priors from LiDAR. Alongside the model, we release a dataset, curated from public sources, consisting of 299 scenes spanning ∼1,244km2 of diverse terrain, paired with 1,130km of human driving. In field trials on a Clearpath Warthog across seven routes at two sites,our method achieves trajectories within 5.5% of human path length and reduces operator interventions by ∼85% compared to local-planner-only autonomy.
Kasi Viswanath, Jason M. Gregory, Shaunak Kolhe +1
J. Mike Walker ’66 Department of Mechanical Engineering, Texas A&M University, College Station, TX 77840, USA · DEVCOM Army Research laboratory, Adelphi, MD, USA
Traversability analysis in off-road settings remains a fundamental challenge for mobile robots. Key difficulties include constructing an accurate and expressive local map from multimodal sensor data, and using the map to design traversability rules that yield desirable navigation behavior. Importantly, this system must be resilient to the limited sensing regime brought about by complex environments and high speeds. In this paper, we present TravSUITE, a traversability system suitable for high-speed navigation in off-road environments. TravSUITE consists of two major components: 1) an efficient voxel mapper that leverages visual foundation models (VFMs) to build a rich geometric-semantic local map from streams of on-board sensor data, and 2) a unified neural network that jointly predicts traversability-relevant quantities in bird's eye view (BEV), including geometry, semantics, speed and cost. Our training strategy is entirely annotation-free and self-supervised, leveraging tasks such as map inpainting and inverse reinforcement learning (IRL) to learn both map representations and traversability. We perform a thorough ablation study and comparison to state-of-the-art approaches, and the results indicate that cost learning and auxiliary inpainting each contribute significantly to planning quality, and their combination is critical for achieving state-of-the-art performance in off-road path planning. We also design a simple risk adaptation mechanism to leverage our method's uncertainty estimates at deploy-time, and demonstrate that a combination of inpainting and risk estimation can result in 80% fewer navigation errors and 5% faster autonomous traversal speeds in real-world hardware experiments.
Samuel Triest, Amirreza Shaban, David D. Fan +2
Robotics Institute, Carnegie Mellon University · Field AI
Accurate, up-to-date representations of road structures are critical for the safe operation of autonomous vehicles. Existing systems rely either on costly, maintenance-heavy high-definition (HD) maps which compromise safety when outdated, or purely sensor-based online mapping which struggles with long-range reliability and occlusion. Systems incorporating map prior information into online mapping seek to overcome drawbacks of both approaches by combining them in some way. We propose 'Driveline To HD Map' (D2HDMap), an online mapping system that injects a lightweight, non-visible driveline prior to guide the estimation of visible road structures such as lane dividers, road boundaries and crosswalks. This prior incurs less effort to create and update compared to full HD map priors used in other approaches. We also show that training with such a prior can improve generalization at inference time when no prior is available. Ablation studies conducted on the nuScenes and Argoverse 2 dataset demonstrate that models trained using a driveline prior largely retain performance even when priors are not available. On a geographically disjoint split, D2HDMap achieves 44.8 mAP, surpassing recent state-of-the-art. Additionally, noise-aware training substantially increases robustness to realistic localization error.
Seojun Shon, Chikao Tsuchiya, Dhaval Bhanderi +3
Nissan North America · Nissan Advanced Technology Center - Silicon Valley, Nissan North America, Santa Clara, CA 95014, USA