@article{Li2026, 
author = {Linheng Li and Weijie Wang and Tiantong Nan and Jing Gan and Xu Qu and Bin Ran},
title = {VTF-Flow: Unified Terrain–Vehicle Kinematic Feasibility Guidance for Generative Trajectory Planning in Unstructured Off-Road Environments},
year = {2026},
journal = {Journal of Intelligent and Connected Vehicles},
keywords = {Unstructured Off-road Environments, Off-road Trajectory Planning, Flow Matching, Vehicle-conditioned Feasibility, Generative Motion Planning, Terrain Traversability},
url = {https://www.sciopen.com/article/10.26599/JICV.2026.9210112},
doi = {10.26599/JICV.2026.9210112},
abstract = {Trajectory planning in unstructured off-road environments must handle weak road boundaries, continuous terrain variation, and multimodal local motion. Although generative planners can represent multiple futures, terrain feasibility is often treated as a static mask or post-generation check, separating it from candidate-induced motion states and trajectory geometry. We propose VTF-Flow, a Vehicle-Conditioned Terrain Feasibility-Guided Flow Planning framework that embeds a unified terrain–vehicle kinematic potential into trajectory generation. It constructs a continuous static potential from LiDAR geometry and semantic labels, conditions it on candidate speed and reliable heading, and adds soft curvature and lateral-acceleration penalties. A bounded feasibility score regularises Flow Matching during training, while an endpoint-preserving trajectory-coordination operator injects the potential gradient into every inference-time integration step. Across 1,909 paired diagnostics on RELLIS-3D sequence 00004, all controlled perturbations reduced relative feasibility; obstacle-directed perturbations achieved a 77.58% win rate, and five-step gradient descent increased feasibility while reducing occupancy exposure. In a sequence-level outer-holdout benchmark of 7,063 planning scenes, VTF-Flow achieved the lowest minADE@8 and unified TVK potential, as well as the highest validation-derived q80 candidate compliance and goal-conditioned compliant coverage at K=8. Relative to matched unguided Flow Matching, it reduced ADE-0, minADE@8, the TVK potential, and curvature violation by 1.31%, 1.07%, 0.99%, and 17.70%, respectively; the principal feasibility metrics improved in all nine seed–held-out-sequence pairs. Continuous in-flow correction therefore improved the availability of low-potential, goal-consistent trajectories while retaining the learned multi-candidate prior.}
}