TruckDrive: Long-Range Autonomous Highway Driving Dataset

Filippo Ghilotti; Edoardo Palladin; Samuel Brucker; Adam Sigal; Mario Bijelic; Felix Heide

cs.CVMar 2, 2026

TruckDrive: Long-Range Autonomous Highway Driving Dataset

Authors: Filippo Ghilotti, Edoardo Palladin, Samuel Brucker, Adam Sigal, Mario Bijelic, Felix Heide

Organizations: Torc Robotics · Princeton University

Abstract

Safe highway autonomy for heavy trucks remains an open and unsolved challenge: due to long braking distances, scene understanding of hundreds of meters is required for anticipatory planning and to allow safe braking margins. However, existing driving datasets primarily cover urban scenes, with perception effectively limited to short ranges of only up to 100 meters. To address this gap, we introduce TruckDrive, a highway-scale multimodal driving dataset, captured with a sensor suite purpose-built for long range sensing: seven long-range FMCW LiDARs measuring range and radial velocity, three high-resolution short-range LiDARs, eleven 8MP surround cameras with varying focal lengths and ten 4D FMCW radars. The dataset offers 475 thousands samples with 165 thousands densely annotated frames for driving perception benchmarking up to 1,000 meters for 2D detection and 400 meters for 3D detection, depth estimation, tracking, planning and end to end driving over 20 seconds sequences at highway speeds. We find that state-of-the-art autonomous driving models do not generalize to ranges beyond 150 meters, with drops between 31% and 99% in 3D perception tasks, exposing a systematic long-range gap that current architectures and training signals cannot close.

Figures & tables

Figure 2: Performance Saturation on Urban Datasets. We plot the performance of 2D and 3D OD, Tracking, Prediction and Depth Estimation of NuScenes [ 6 ] and Kitti [ 3 ] leader boards across the years and observe a saturation of these benchmarks.

Dataset	LiDARs	Radars	Cameras	Localization	Sensor	Synced	Manually	Effective
Dataset	LiDARs	Radars	Cameras	Localization	Count	Samples	Annotated	Range
KITTI [ 22 ]	1x 64-beam	-	2x RGB, 2x grayscale	GPS, IMU	7	216k	15k	[0, +70]
SemanticKITTI [ 2 ]	1x 64-beam	-	2x RGB, 2x grayscale	GPS, IMU	7	43k	43k	[-80, +80]
ApolloScape [ 58 ]	2x	-	2x RGB, 2x grayscale	GPS, IMU	8	143k	20k	[-100, +100]
A2D2 [ 23 ]	5x 16-beam	-	6x	GPS, IMU	13	392k	40k	[-100, +100]
H3D [ 49 ]	1x 64-beam	-	3x	GPS, IMU	6	27k	27k	[-100, +100]

Table 1: TruckDrive Benchmark Comparison. Cross-dataset summary of sensors, synced samples and useful ranges. TruckDrive couples

7

long range and

3

short range LiDARs with

10

automotive radars,

9

wide/medium field of view cameras and

1-3

long-focal-length wide-baseline stereo cameras. It offers

165

thousands annotated samples and additional

310

thousands unlabeled samples and extends the effective perception range to [

-400

+400

] meters, focusing on highway long-range scenarios to stress perception capabilities beyond conventional benchmarks. ∗ NuPlan [ 7 ] provides auto-labeled annotations.

Figure 3: Sensors Position and FoV. Sensor position (top) and the nominal instrumented horizontal field of view (bottom) of, from left to right, radars, LiDARs and cameras, highlighting the unprecedented ranges at which they can operate.

Panel of Figure 4 — Figure 4: Dataset Analysis. Our dataset comprises an unprecedented density of instance objects at ranges (greater than $200$ meters) yet to be explored in publicly available datasets (a,b), as well as driving speeds $5$ times higher (c) and sequences with traveled length up to $8$ times longer (d) than existing benchmarks.

Figure 5: Driving Tasks and Challenges. We report qualitative results of the best baselines across planning, 2D/3D object detection, depth estimation and scene reconstruction. Even when trained on TruckDrive, existing methods struggle in the long-range, high-speed regime. Planning modules exhibit conservative behavior due to low-speed assumptions. Grid-based BEV models degrade perception as large spatial coverage demands heavy downsampling, erasing safety-critical details such as small debris or lost cargo [ 40 ] , while depth methods struggle beyond 200m and in sky regions [ 25 ] , revealing limited distance awareness and motivating architectures for highway-scale perception.

Method	mAP $\uparrow$	mAP 50 $\uparrow$	mAP 75 $\uparrow$	mAP SR $\uparrow$	mAP MR $\uparrow$	mAP LR $\uparrow$	mAP UR $\uparrow$
DETR [ 8 ]	12.70%	23.90%	12.20%	41.20%	24.70%	8.90%	1.00%
ViTDet [ 36 ]	27.30%	37.60%	30.80%	58.30%	51.80%	33.90%	3.30%
YOLO11x [ 32 ]	28.90%	39.00%	31.60%	36.30%	29.40%	8.20%	2.00%
DINO [ 66 ]	37.80%	54.20%	40.30%	63.90%	54.60%	43.20%	15.30%

Table 3: 2D Object Detection Results. We follow CoCo [ 39 ] and report mean average precision (mAP) at

0.50

IoU, mAP at

0.75

IoU, and mAP at short (

0-50

m, SR), medium (

50-150

m, MR), long (

150-250

m, LR), and ultra-long-range (

250

+, UR).

Method	Mode	mAP $\uparrow$	mAP $\uparrow$	mAP $\uparrow$	mAP $\uparrow$
Method	Mode	Full	SR	MR	LR
Far3D [ 31 ]	C	14.04%	35.54%	11.07%	0.33%
TransFusion-L [ 1 ]	L	25.24%	30.12%	22.25%	22.25%
BEVFusion [ 40 ]	L+C	26.45%	32.32%	22.77%	22.69%

Table 4: 3D Object Detection Results . We report mAP for

3

baselines using a single or a combination of LiDAR (L) and Camera (C) data, divided into short (

0-50

m, SR), medium (

50-150

m, MR), long (

150-250

m, LR) and full detection ranges.

Method	Mode	AMOTA $\uparrow$	AMOTP $\downarrow$	Recall $\uparrow$
MUTR3D [ 67 ]	Query	6.1%	79.0%	11.4%
Immortal Tracker † [ 59 ]	3D Box	12.8%	77.2%	20.7%
CenterPoint † [ 65 ]	3D Box	13.0%	76.9%	21.5%

Table 5: 3D Multi Object Tracking Results . We report AMOTA, AMOTP and Recall for a query based and two LiDAR based methods. † uses [ 40 ] inference detections.

Table 6: Depth Estimation Results . We report performances for surround (a), stereo (b) and monocular (c) depth estimation tasks. Each method is evaluated with standard accuracy and error metrics at short (0-50m, SR), medium (50-150m, MR), long (150-250m, LR) and ultra (250-1000m, UR) range bins.

History Horizon	Method	Modality	1s		3s
History Horizon	Method	Modality	CD $\downarrow$	L1 (m) $\downarrow$	CD $\downarrow$	L1 (m) $\downarrow$
1s	4DOcc [ 34 ]	L	18.93	4.69	-	-
	ViDAR [ 64 ]	C	58.23	20.21	51.72	20.69
	LRS4Fusion [ 47 ]	L + C	15.82	3.31	39.03	3.82
3s	4DOcc [ 34 ]	L	23.58	3.00	47.81	4.29
	ViDAR [ 64 ]	C	57.28	20.14	56.20	20.53

Table 7: LiDAR Forecasting Results . We evaluate single and multi-modal state of the arts methods with Chamfer Distances (CD) of

1

and

3

seconds and L1 error.

Method	SR		MR		LR		FULL
Method	F1 $\uparrow$	$\mathbf{IoU\uparrow}$	F1 $\uparrow$	$\mathbf{IoU\uparrow}$	F1 $\uparrow$	$\mathbf{IoU\uparrow}$	F1 $\uparrow$	$\mathbf{IoU\uparrow}$
4DMOS ‡ [ 45 ]	25.9	18.5	8.4	6.1	0.6	0.4	24.4	16.7
4DMOS [ 45 ]	47.3	32.1	22.7	15.4	8.3	5.6	31.8	21.6

Table 8: 3D Moving Object Segmentation Results . ‡ indicates results from the public KITTI [ 22 ] checkpoint.

Method	Representation	PSNR $\uparrow$	SSIM $\uparrow$
Dyn. Nerfacto [ 55 ]	NeRF	26.2870	0.8653
HUGS [ 70 ]	3DGS	29.2675	0.8858
OmniRe [ 11 ]	3DGS	33.8244	0.9515

Table 9: 3D Reconstruction Quality Results. We report PSNR and SSIM for a NeRF and two 3D Gaussian Splatting methods.

Method	$L2$ (m) $\downarrow$
Method	1 step	2 step	3 step	4 step	5 step	6 step	Avg.
UniAD [ 29 ]	0.57	1.13	1.71	2.30	2.88	3.42	2.00

Table 10: E2E Planning. We train UniAD [ 29 ] on our long range setup and evaluate L2 error for all predicted time intervals.

Autonomous Driving Datasets

Autonomous driving datasets collect and organize driving-scene data to train and evaluate perception, prediction, planning, and safety systems.

See all 136 papers

Telescope: Learnable Hyperbolic Foveation for Ultra-Long-Range Object Detection
Apr 7, 2026Parker Ewen, Dmitriy Rivkin, Mario Bijelic +1Foveated Vision Object Detection
Torc Robotics · Princeton University
A2RL V\textsubscript{max}: The A2RL autonomous racing dataset for long-range, high-speed perception and multi-vehicle interaction
Jul 20, 2026Marvin Klemp, Dominic Ebner, Cornelius Schröder +15Autonomous Driving Perception Autonomous Racing
Institute of Measurement and Control Systems, Karlsruhe Institute of Technology · Code 19 Racing · Institute of Automotive Technology, Technical University of Munich +8
The Road Ahead in Autonomous Driving: The KITScenes Multimodal Dataset
Jun 1, 2026Richard Schwarzkopf, Fabian Immel, Alexander Blumberg +21Autonomous Driving Datasets Autonomous Driving Benchmarks
FZI Research Center for Information Technology · Karlsruhe Institute of Technology · University Charles III of Madrid +1

Apr 7, 2026J/K move · Enter open · S save

Telescope: Learnable Hyperbolic Foveation for Ultra-Long-Range Object Detection

Parker Ewen, Dmitriy Rivkin, Mario Bijelic, Felix Heide

Torc Robotics · Princeton University

Foveated Vision Object Detection Autonomous Driving Perception Small Object Detection

Open paper AskPDF arXiv

Autonomous highway driving, especially for long-haul heavy trucks, requires detecting objects at long ranges beyond 500 meters to satisfy braking distance requirements at high speeds. At long distances, vehicles and other critical objects occupy only a few pixels in high-resolution images, causing state-of-the-art object detectors to fail. This challenge is compounded by the limited effective range of commercially available LiDAR sensors, which fall short of ultra-long range thresholds because of quadratic loss of resolution with distance, making image-based detection the most practically scalable solution given commercially available sensor constraints. We introduce Telescope, a two-stage detection model designed for ultra-long range autonomous driving. Alongside a powerful detection backbone, this model contains a novel re-sampling layer and image transformation to address the fundamental challenges of detecting small, distant objects. Telescope achieves

76\%

relative improvement in mAP in ultra-long range detection compared to state-of-the-art methods (improving from an absolute mAP of 0.185 to 0.326 at distances beyond 250 meters), requires minimal computational overhead, and maintains strong performance across all detection ranges.

Abstract

Autonomous Driving Datasets

Explore similar work

Telescope: Learnable Hyperbolic Foveation for Ultra-Long-Range Object Detection

A2RL V\textsubscript{max}: The A2RL autonomous racing dataset for long-range, high-speed perception and multi-vehicle interaction

The Road Ahead in Autonomous Driving: The KITScenes Multimodal Dataset