Mind the Hitch: Dynamic Calibration and Articulated Perception for Autonomous Trucks

Morui Zhu; Yongqi Zhu; Song Fu; Qing Yang

cs.CVMar 24, 2026

Mind the Hitch: Dynamic Calibration and Articulated Perception for Autonomous Trucks

Authors: Morui Zhu, Yongqi Zhu, Song Fu, Qing Yang

Organizations: University of North Texas

Abstract

Autonomous trucking poses unique challenges due to articulated tractor-trailer geometry, and time-varying sensor poses caused by the fifth-wheel joint and trailer flex. Existing perception and calibration methods assume static baselines or rely on high-parallax and texture-rich scenes, limiting their reliability under real-world settings. We propose dCAP (dynamic Calibration and Articulated Perception), a vision-based framework that continuously estimates the 6-DoF (degree of freedom) relative pose between tractor and trailer cameras. dCAP employs a transformer with cross-view and temporal attention to robustly aggregate spatial cues while maintaining temporal consistency, enabling accurate perception under rapid articulation and occlusion. Integrated with BEVFormer, dCAP improves 3D object detection by replacing static calibration with dynamically predicted extrinsics. To facilitate evaluation, we introduce STT4AT, a CARLA-based benchmark simulating semi-trailer trucks with synchronized multi-sensor suites and time-varying inter-rig geometry across diverse environments. Experiments demonstrate that dCAP achieves stable, accurate perception while addressing the limitations of static calibration in autonomous trucking. The dataset, development kit, and source code will be publicly released.

Figures & tables

Figure 1: Overview of dCAP. We perform online 6-DoF articulated pose estimation for tractor–trailer systems and enable articulated-aware perception. Unlike traditional SfM methods (e.g., COLMAP), dCAP succeeds without requiring a valid static initialization pair.

Sensor	Details
6x Camera	RGB, 1600 $\times$ 900 resolution, 110 ∘ FOV
1x LiDAR	128 channels, 3.5M points per second,
	200 m capturing range, -20 ∘ to 20 ∘ vertical
	FOV, $\pm$ 2 cm error
2x GPS & 2x IMU	20 mm positional error, 2 ∘ heading error

Table 1: Sensor specifications.

Figure 2: Example from the STT4AT dataset showing six synchronized camera views, the LiDAR point cloud with annotated agents, a BEV illustration of trailer articulation.

Figure 3: Distribution of annotated frames in the STT4AT dataset. The left chart separates straight and turning maneuvers, while the right details the composition within turning scenarios.

Figure 4: Overview of the proposed architecture. Multi-view images at time $t$ are encoded by a frozen VGGT backbone into camera tokens, while the trainable decoder comprises (a) Camera Temporal Self-Attention (CTA) for fusing the historical token $T_{t-1}$ with the current query $Q$ , (b) Camera Cross-Attention (CCA) for attending $Q$ to encoder tokens $\{T_{i}\}_{i=1}^{6}$ , (c) an AdaLN-modulated refinement stack with residual Add&Norm applied $L$ times.

Method	$\Delta_{T}\downarrow$	$\Delta_{x}\downarrow$	$\Delta_{y}\downarrow$	$\Delta_{z}\downarrow$	RRA $\downarrow$
Static Calibration	1.284	0.210	1.120	0.356	0.148
COLMAP [ 23 ] †	-	-	-	-	-
VGGT [ 29 ]	6.040	2.761	3.082	3.634	0.309
DUSt3R [ 31 ]	8.625	4.664	5.080	2.953	0.578
GNSS-IMU (Kalman Filter)	1.379	0.309	1.116	0.431	0.129
dCAP (w/o CCA, w/o CTA)	0.632	0.076	0.600	0.087	0.073

Table 2: Quantitative results of trailer camera pose prediction under different methods. Note that † fails to reconstruct due to the lack of a valid initial image pair.

Method	AP $\uparrow$	NDS $\uparrow$	ATE $\downarrow$	ASE $\downarrow$	AOE $\downarrow$	AVE $\downarrow$	AAE $\downarrow$	[email protected] $\uparrow$	[email protected] $\uparrow$	[email protected] $\uparrow$	[email protected] $\uparrow$
Static Calibration	0.058	0.033	0.734	0.176	0.153	2.419	0.237	0.0000	0.0188	0.0728	0.1417
COLMAP [ 23 ] †	-	-	-	-	-	-	-	-	-	-	-
Tractor only	0.049	0.032	0.705	0.183	0.198	2.572	0.267	0.0000	0.0159	0.0623	0.1173
VGGT [ 29 ]	0.033	0.031	0.671	0.181	0.202	2.619	0.251	0.0000	0.0115	0.0442	0.0783
DUSt3R [ 31 ]	0.034	0.031	0.711	0.182	0.219	2.682	0.256	0.0000	0.0092	0.0422	0.0839
dCAP (w/o CCA, w/o CTA)	0.084	0.034	0.710	0.172	0.132	2.328	0.222	0.0007	0.0386	0.1105	0.1863

Table 3: Quantitative results of 3D object detection under different attention configurations and baselines. Note that † fails to reconstruct due to the lack of a valid initial image pair.

CCA	CTA	$\Delta_{T}\downarrow$	$\Delta_{x}\downarrow$	$\Delta_{y}\downarrow$	$\Delta_{z}\downarrow$	$RRA\downarrow$
✗	✗	0.926	0.037	0.913	0.079	0.071
✓	✗	0.517	0.046	0.501	0.051	0.051
✗	✓	0.459	0.105	0.430	0.052	0.058

Table 4: Trailer pose estimation under the Straight scenario.

CCA	CTA	$\Delta_{T}\downarrow$	$\Delta_{x}\downarrow$	$\Delta_{y}\downarrow$	$\Delta_{z}\downarrow$	$RRA\downarrow$
✗	✗	0.850	0.136	0.803	0.097	0.095
✓	✗	0.675	0.119	0.634	0.094	0.061
✗	✓	0.475	0.168	0.398	0.100	0.066

Table 5: Trailer pose estimation under the Roundabout scenario.

CCA	CTA	$\Delta_{T}\downarrow$	$\Delta_{x}\downarrow$	$\Delta_{y}\downarrow$	$\Delta_{z}\downarrow$	$RRA\downarrow$
✗	✗	1.212	0.159	1.175	0.154	0.112
✓	✗	1.117	0.188	1.083	0.119	0.091
✗	✓	0.706	0.202	0.642	0.097	0.100

Table 6: Trailer pose estimation under the U-turn scenario.

CCA	CTA	$\Delta_{T}\downarrow$	$\Delta_{x}\downarrow$	$\Delta_{y}\downarrow$	$\Delta_{z}\downarrow$	$RRA\downarrow$
✗	✗	0.520	0.081	0.459	0.140	0.066
✓	✗	0.361	0.069	0.286	0.118	0.037
✗	✓	0.423	0.140	0.325	0.129	0.058

Table 7: Trailer pose estimation under the Multi-Turn scenario.

CCA	CTA	mAP $\uparrow$	ATE $\downarrow$	ASE $\downarrow$	AOE $\downarrow$	AVE $\downarrow$	AAE $\downarrow$
✗	✗	0.0481	0.9598	0.9175	0.8972	1.0135	0.9104
✓	✗	0.0549	0.9561	0.9176	0.8964	1.0064	0.9091
✗	✓	0.0497	0.9632	0.9173	0.8966	1.0051	0.9103
GT (upper bound)		0.0686	0.9414	0.9174	0.8960	1.0035	0.9096

Table 8: Detection performance under the Straight scenario.

CCA	CTA	mAP $\uparrow$	ATE $\downarrow$	ASE $\downarrow$	AOE $\downarrow$	AVE $\downarrow$	AAE $\downarrow$
✗	✗	0.0290	0.9800	0.9177	0.9402	1.1919	0.9255
✓	✗	0.0324	0.9770	0.9175	0.9326	1.1998	0.9206
✗	✓	0.0397	0.9730	0.9182	0.9291	1.2058	0.9152
GT (upper bound)		0.0507	0.9564	0.9180	0.9229	1.1704	0.9203

Table 9: Detection performance under the Roundabout scenario.

CCA	CTA	mAP $\uparrow$	ATE $\downarrow$	ASE $\downarrow$	AOE $\downarrow$	AVE $\downarrow$	AAE $\downarrow$
✗	✗	0.0416	0.9760	0.9183	0.8996	1.1722	0.8925
✓	✗	0.0464	0.9689	0.9180	0.8988	1.1518	0.8939
✗	✓	0.0482	0.9683	0.9179	0.8981	1.1534	0.8905
GT (upper bound)		0.0663	0.9424	0.9176	0.8984	1.1604	0.8901

Table 10: Detection performance under the U-turn scenario.

CCA	CTA	mAP $\uparrow$	ATE $\downarrow$	ASE $\downarrow$	AOE $\downarrow$	AVE $\downarrow$	AAE $\downarrow$
✗	✗	0.0437	0.9641	0.9166	0.8988	1.1474	0.8770
✓	✗	0.0496	0.9627	0.9166	0.8985	1.1482	0.8753
✗	✓	0.0473	0.9606	0.9168	0.8999	1.1593	0.8783
GT (upper bound)		0.0553	0.9509	0.9163	0.8985	1.1435	0.8758

Table 11: Detection performance under the Multi-Turn scenario.

Appendix figures & tables7 assets

Supplementary material from the paper’s appendix.

Appendix

Figure 5: Representative articulated-driving scenarios in the STT4AT dataset. Each example shows the tractor–trailer configuration and trailer-mounted camera poses under typical high-articulation maneuvers.

Figure 6: Placement of the sensors and their corresponding coordinate systems from a top-down view perspective

Module	#Params (M)	Trainable
VGGT backbone	909.1	✗
Camera Cross-Attention (CCA)	16.8	✓
Camera Temporal Self-Attention (CTA)	16.8	✓
AdaLN-guided pose refinement module	216.2	✓

Appendix

Table 12: Number of parameters for core modules in dCAP.

Figure 7: Qualitative results comparison between different attention modules under the straight scenario.

Figure 8: Qualitative results comparison between different attention modules under the roundabout scenario.

Figure 9: Qualitative results comparison between different attention modules under the u-turn scenario.

Figure 10: Qualitative results comparison between different attention modules under the multi-turn scenario.

Autonomous Driving

Autonomous driving develops vehicles that perceive road scenes, predict interactions, and plan and control safe movement with limited or no human input.

See all 536 papers

Hybrid Machine Learning for Articulation Angle Estimation of Truck-Semitrailer Combinations
Jul 18, 2026Qixuan Zhang, Jonas Boettcher, Simon F. G. Ehlers +1Autonomous Driving Perception Robot State Estimation
ZF CV Systems GmbH, D-30453 Hannover, Germany · Leibniz University Hannover, Institute of Mechatronic Systems, D-30823 Garbsen, Germany
TRIG: Trajectory-Rig Decoupled Metric Geometry Learning
Jul 7, 2026Lizhou Liao, Wentao Xu, Handong Wang +4Multi-View Geometry Multi-View 3D Reconstruction
Carizon · ShanghaiTech University
Parking Assistance for Trailer-Truck Transport Vehicles Using Sensor Fusion and Motion Planning
May 4, 2026George Alenchery, Thomas Jeske, Tova Quinones +4Nonlinear MPC Autonomous Parking
Software Engineering Department Florida Gulf Coast University

Jul 18, 2026J/K move · Enter open · S save

Hybrid Machine Learning for Articulation Angle Estimation of Truck-Semitrailer Combinations

Qixuan Zhang, Jonas Boettcher, Simon F. G. Ehlers, Marvin Stuede

ZF CV Systems GmbH, D-30453 Hannover, Germany · Leibniz University Hannover, Institute of Mechatronic Systems, D-30823 Garbsen, Germany

Autonomous Driving Perception Robot State Estimation

Open paper AskPDF arXiv

Accurate articulation angle estimation of trucks with trailers is critical for autonomous driving and advanced driver assistance system (ADAS). Existing methods either require manual initialization, additional sensors, or prior knowledge and signals from trailers, or they lack real-world validation, limiting practical deployment. This paper presents multiple learning-based models to directly estimate articulation angles from visual and kinematic inputs, eliminating the need for dedicated driving maneuvers for initialization, bounding box annotations, trailer-mounted sensor signals, or prior knowledge of trailer parameters. Two learning-based models are integrated with a kinematic model within an extended Kalman filter (EKF) framework, and an adaptive weighting scheme based on uncertainty quantification is applied for measurements involving visual input. Extensive real-world experiments with different trailer types demonstrate the approaches' robustness and generalization under out-of-domain conditions, including new trailers, varying colors, and lighting conditions. Results show that the hybrid method achieves accurate and reliable articulation angle estimation while maintaining reduced implementation requirements and practical deployment advantages.

Abstract

Autonomous Driving

Explore similar work

Hybrid Machine Learning for Articulation Angle Estimation of Truck-Semitrailer Combinations

TRIG: Trajectory-Rig Decoupled Metric Geometry Learning

Parking Assistance for Trailer-Truck Transport Vehicles Using Sensor Fusion and Motion Planning