TravSUITE: Traversability via Self-Supervised, Uncertainty-Aware IRL and Terrain Estimation

Samuel Triest; Amirreza Shaban; David D. Fan; Wenshan Wang; Sebastian Scherer

cs.ROOct 23, 2026

TravSUITE: Traversability via Self-Supervised, Uncertainty-Aware IRL and Terrain Estimation

Authors: Samuel Triest, Amirreza Shaban, David D. Fan, Wenshan Wang, Sebastian Scherer

Organizations: Robotics Institute, Carnegie Mellon University · Field AI

Abstract

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.

Figures & tables

Fig. 2: An overview of our supervision. TravSUITE relies entirely on supervision from hindsight, including trajectory and map information. (Blue=derived from current observations, Red=derived from privileged information)

Feature	Modality	Rep.	Description
Terrain	Geom.	BEV	Elevation of the terrain surface
Slope	Geom.	BEV	Slope of the terrain surface
Diff	Geom.	BEV	Max elevation over the terrain
min/max $z$	Geom.	Voxel	The min/max height in each voxel
Seg Logit x16	Seg.	Both	Class score from Talk2Dino [ 7 ]
VFM x32	VFM	Both	VFM feature from Radiov3 [ 50 ]

TABLE I: List of Mapping features

Fig. 3: An overview of our network architecture. The TravSUITE network first starts with a ResNet-style encoder (either 2d or 3d depending on the input modality), followed by BEV-space inpainting and parallel BEV decoder FCN heads. Since the individual prediction heads are relatively lightweight, the backbone and encoder must handle the majority of the feature extraction.

Fig. 4: A diagram of the cost architecture and uncertainty mechanism. (Top-Left) A sample costmap at CVaR=0. (Bottom-Left) The same costmap at CVaR=0.8. (Right) A zoom-in of the cost distribution for two select cells. The red cell has a low-entropy cost distribution and thus stays at a similar value for all CVaR values. The cyan cell has a high-entropy cost distribution and has its cost inflated significantly under higher CVaR.

Inpainting	Dataset	MHD $\downarrow$	Speed Prob. $\uparrow$
All	Test 1	1.3936	0.1159
Geometry	Test 1	1.4138	0.1054
VFM	Test 1	1.5400	0.1162
Segmentation	Test 1	1.4877	0.1120
None	Test 1	1.7744	0.1026
All	Test 2	1.3988	0.1258

TABLE II: Effect of inpainting on planning tasks

Input	Dataset	MHD $\downarrow$	Speed Prob. $\uparrow$
Voxel	Test 1	1.3936	0.1159
BEV	Test 1	1.4440	0.1184
Voxel no agg.	Test 1	1.5061	0.1109
Voxel	Test 2	1.3988	0.1258
BEV	Test 2	1.4373	0.1311
Voxel no agg.	Test 2	1.4452	0.1031

TABLE III: Effect of input representation on planning tasks

Input	Dataset	Geom $\downarrow$	VFM $\downarrow$	Seg $\downarrow$
Voxel	Test 1	0.7253	1.2324	0.3087
BEV	Test 1	0.7137	1.1450	0.3019
Voxel No Agg.	Test 1	1.5952	1.9533	0.5064
Voxel	Test 2	0.5221	1.2389	0.4279
BEV	Test 2	0.4704	1.1313	0.4019
Voxel No Agg.	Test 2	1.6326	2.0331	0.6033

TABLE IV: Effect of input representation on inpainting tasks

Approach	Dataset	MHD $\downarrow$	Speed Prob. $\uparrow$
Parallel Trav.	Test 1	1.3936	0.1159
Cascaded Trav.	Test 1	1.3043	0.1076
TerrainNet Cfn.	Test 1	1.4553	N/A
FCN no Inpaint	Test 1	1.5246	0.0967
Parallel Trav.	Test 2	1.3988	0.1258
Cascaded Trav.	Test 2	1.3845	0.1156

TABLE V: Comparison to similar methods

Fig. 5: Two scenarios from the test set. Cost and speed maps are cropped and enlarged to highlight regions of interest. (Top row): The robot must navigate through a path obscured by dense vegetation (cyan arrow). The Voxel+Inpaint network does the best at assigning low cost to the trail and high cost to the surrounding unobserved space, which is likely vegetation. (Bottom row): The robot is traveling at high speed on an open trail. Due to the speed, LiDAR measurements are relatively sparse. The inpainting method is able to fill in the missing trail features and keep a consistent speed map.

Method	Col.	Int.	Avg. Speed	Top Speed
Parallel Trav.	2	2	4.79m/s	7.66m/s
Terrainnet Cfn.	1	1	5.23m/s	7.72m/s
FCN no Inpaint	1	4	4.33m/s	6.84m/s
Parallel Trav. + CVaR	0	1	4.49m/s	7.16m/s

TABLE VI: Hardware Metrics

Fig. 6: GPS map of our navigation course. The course totaled about 1.4km in length and contained a number of challenges including snow, vegetation and man-made obstacles. In order to successfully navigate the course, the vehicle must be able to avoid OOD obstacles, drive through tall vegetation, and intelligently select traversal speed based on the terrain surface.

Fig. 7: Qualitative figure from the navigation trial. (a, b) After reaching the current waypoint, the ATV must navigate to another waypoint about $150m$ to its left. (c, d) Without inpainting, the FCN baseline underestimates the cost of the vegetation out of the camera view, resulting in an intervention. (e, f) The inpainting baseline is able to successfully estimate the cost of the vegetation, thus encouraging the local planner to take a longer path on the trail.

Fig. 8: Qualitative figure from the navigation trial. The robot was tasked with navigating around the set of cones in the trail. (Left) Inpaint + IRL without CVaR plans through the cones. (Center) The TerrainNet cfn does not sufficiently distinguish between the cones and the tall vegetation to its side. (Right) IRL with CVaR assigns enough cost to the cones for the vehicle to plan around into the grass.

Appendix figures & tables8 assets

Supplementary material from the paper’s appendix.

Appendix

Fig. 9: Plot of the CVaR value during the Parallel Trav. + CVaR autonomy run.

Fig. 10: Qualitative example comparing voxel inputs versus BEV. (Yellow box): There is a low-hanging canopy over the trail, which the demonstration trajectory (yellow arrow) goes under. The voxel network is correctly able to predict the trail as low cost, while the BEV network misclassifies the trail as an obstacle due to the limited vertical representativeness of the BEV representation.

	Image Feature Extraction	Feature Lifting	Mapping	BEV Feature	Inference
Voxel	26.3/25.6/27.0	2.8/2.6/3.9	8.5/8.1/11.2	12.3/12.0/13.6 ∗	31.1/74.7/103.5
BEV	26.4/25.7/27.9	2.8/2.7/4.9	3.5/3.2/7.3	15.8/14.6/19.4	25.9/27.2/32.0

Appendix

TABLE VII: Timing speeds (mean/min/max in ms)

Fig. 11: A simple 2d example to highlight the intuition behind the voxel hindsight generation algorithm. (Blue) The state $x_{t}$ and corresponding mapping volume $\mathcal{M}_{t}$ . (Cyan) The state $x_{t^{\prime}}$ and corresponding mapping volume $\mathcal{M}_{t^{\prime}}$ . Note that $\mathcal{M}_{t^{\prime}}$ intersects with $\mathcal{M}_{t}$ and $\mathcal{M}_{t^{\prime}+1}$ does not. Since $\mathcal{M}_{t^{\prime}}$ intersects with $\mathcal{M}_{t}$ , it must be the case that a volume with triple the side length, centered at $x_{t}^{\prime}$ (i.e. $\mathcal{M}^{\prime}_{t^{\prime}}$ ) must fully contain $\mathcal{M}_{t}$ .

Parameter	Value
Mapping Origin ( $\mathcal{M}_{o}$ )	$-50m\times-50m\times-10m$
Mapping Length ( $\mathcal{M}_{l}$ )	$100m\times 100m\times 20m$
Mapping Resolution ( $\mathcal{M}_{r}$ )	$0.4m\times 0.4m\times 0.1m$
Mapping Cell Counts ( $\mathcal{M}_{n}$ )	$250\times 250\times 200$
Input Image Shape	$848\times 448\times 3$

Appendix

TABLE VIII: Local Mapping Parameters

Parameter	Value
Voxel Encoder	3D ResNet Block $\times 4$
BEV Encoder	ResNet Block $\times 4$
Backbone	UNet Block $\times 3$
VFM Decoder	ResNet Block $\times 2$
Seg. Decoder	ResNet Block $\times 2$
Geom. Decoder	ResNet Block $\times 2$

Appendix

TABLE IX: Network Parameters

Parameter	Value
Goal Weight ( $K_{2}$ , Eq. 3)	5.0
Vehicle Footprint	$4m\times 2m$
Behavior Weight ( $\lambda_{1}$ , Eq. 4)	1.0
Inpaint Weight ( $\lambda_{2}$ , Eq. 4)	1.0
Geom Inpaint Weight ( $\lambda_{geom}$ , Eq. 9)	1.0
VFM Inpaint Weight ( $\lambda_{vfm}$ , Eq. 9)	1.0

Appendix

TABLE X: Training Parameters

Parameter	Value
CVaR min/max	$[0.0,0.8]$
CVaR vel min/max ( $v_{low},v_{high}$ , Eq. 8)	$[2,5]m/s$
CVaR step ( $\epsilon$ , Eq. 8)	0.02
Lat Acc. Max	$5m/s^{2}$

Appendix

TABLE XI: Hardware Experiment Parameters

Robot Traversability Estimation

See all 27 papers

Sep 22, 2026

FLINT: Fast Lightweight Inference for Traversability

Navigation in off-road conditions is challenging due to the lack of structure. There is no fixed vocabulary for what is traversable. The traversability depends on both the environment and the embodiment's dynamics. Neither of these two variables can be hand-labeled at scale. Thus, traversability has to be learned by the embodiment's own experience. Modern platforms tend to use multiple sensors to estimate traversability and navigate: RGBD cameras, lidar, radar, IMU, with computationally intensive platforms to run inference on neural networks. Against this trend, we propose FLINT, a lightweight traversability estimator: a 21.6M-parameter backbone, 38\times smaller than a comparable foundation-model backbone, that scores higher on held-out terrain probes and runs at 14.7 FPS on CPU alone using a RGB camera has the only sensor. Despite that gap in scale, FLINT produces a cheaper, more accurate costmap than a deployed foundation-model system (WildOS) on 23 of 24 replayed field logs. We compare different self-supervised learning signals and deploy the resulting models on a real platform in closed-loop field trials: the best self-supervised head reaches 99% autonomy over the route, outperforming a human-label-trained baseline deployed live on the same course. Our results show that heavy sensing and computing are not necessary for traversability estimation.

William Bonilla, Maxime Boisvert, David-Alexandre Poissant +2

School of Computer Science, McGill University, Montreal, QC, Canada · Centre de Technologies Avanc´ees (CTA), Sherbrooke, QC, Canada · Universit´e de Sherbrooke, Sherbrooke, QC, Canada

Robot Traversability Estimation Off-Road Robot Navigation Visual Navigation Embodied Navigation

Ask PDF

Jul 26, 2026

Learning Traversability for Long Horizon Off-Road Navigation

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

\sim\!1{,}244\,\mathrm{km}^{2}

of diverse terrain, paired with

1{,}130\,\mathrm{km}

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

\sim\!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

Off-Road Robot Navigation Geospatial Representation Learning Robot Navigation Safe Robot Navigation Robot Traversability Estimation

Ask PDF

Jul 20, 2026

Distilling Global Traversability Priors for Image-based Affordance Prediction in Off-road Environments

Standard methods for autonomous navigation in unstructured terrain are prone to myopic behaviors in long-horizon scenarios. The use of metric maps built from LiDAR or cameras provides necessary local geometry and semantic information but is strictly limited by depth sensing range. By discarding data beyond the mapping horizon robots suffer from suboptimal, short-sighted decisions. To recover this lost information, we focus on extracting long-range traversability-aware frontiers directly from first-person-view (FPV) images. By leveraging satellite imagery, we compute the set of feasible navigation paths for a dataset of image/pose pairs and use them to supervise our network, reducing the need for extensive human demonstration data. We demonstrate that this approach improves performance in long-range off-road navigation over existing methods by more than 10% in various offline benchmarks and reduces the number of human interventions incurred in a set of real-world experiments. More details can be found at https://theairlab.org/ss_frontiers_iros .

Matthew Sivaprakasam, Samuel Triest, Micah Nye +5

Robotics Institute, Carnegie Mellon University · Field AI

Off-Road Robot Navigation Embodied Navigation Robot Navigation Robot Traversability Estimation

Ask PDF

Abstract

Robot Traversability Estimation

Explore similar work

FLINT: Fast Lightweight Inference for Traversability

Learning Traversability for Long Horizon Off-Road Navigation

Distilling Global Traversability Priors for Image-based Affordance Prediction in Off-road Environments