GDP: Enhancing End-To-End Autonomous Driving with Goal-Driven Planner
Qiming Zhang, Yue Zhao, Yujian Wang, Wei Wang, Zetong Yang, Wei Xu, Yin Zhou, Jun Ma
AI summary
Problem
Existing end-to-end autonomous driving frameworks rely on vague semantic commands that lack geometric precision and explicit goal locations, which hinders long-term decision-making and causes planning ambiguities in complex traffic scenarios.
Approach
The authors propose GDP, a plug-and-play module that predicts a scene-aware long-term route to a navigation goal and uses it to structurally refine the near-term planning trajectory through dual interaction with map and agent features.
Key results
- Predicts scene-aware long-term routes from sparse goal points and map features
- Refines near-term trajectories via a dual-branch interaction module with route and scene features
- Reduces average L2 error by 27.4% and collision rate by 25.0% when integrated with UniAD on nuScenes
- Improves planning performance and safety metrics across multiple E2E frameworks on open-loop and non-reactive simulations
Why it matters
It provides a simple, architecture-agnostic way to bridge the gap between high-level navigation goals and low-level trajectory planning, making existing end-to-end autonomous driving systems safer and more reliable for real-world deployment.
Abstract
End-to-end (E2E) autonomous driving has emerged as a promising paradigm with the pervasive power of model architectures and the availability of large-scale driv- ing datasets. Despite tremendous efforts in recent research, most E2E driving frameworks rely on rather general driving commands, such as “Go Straight” or “Turn Left”, which fail to encapsulate the complexities of nuanced driving behaviors and lead to possible semantic ambiguities. Furthermore, such commands are not adequately translated into specific goal locations, which severely limits the planner’s capacity to make informed, long-term decisions. This limitation hinders the in- tegration of near-term trajectory planning with long-term goal achievement. To tackle these challenges, we propose the Goal- Driven Planner (GDP), accommodating an appealing plug-and- play feature, which particularly leverages explicit goal points and incorporates two complementary learning objectives: (i) predicting a scene-aware long-term route to the goal, and (ii) refining the near-term trajectory through interaction with the long-term routing. When integrated into off-the-shelf E2E autonomous driving frameworks such as UniAD, VAD-Tiny, and DiffusionDrive, GDP improves trajectory quality and safety across most open-loop and non-reactive simulation metrics. By explicitly modeling a goal-driven route and using it as structured guidance for trajectory refinement, GDP provides a complementary planning signal that enhances long-term goal alignment without modifying the underlying E2E architectures.