SAH-Drive: A Scenario-Aware Hybrid Planner for Closed-Loop Vehicle Trajectory Generation

Yuqi Fan, Zhiyong Cui, Zhenning Li, Yilong Ren, Haiyang Yu

Abstract

Autonomous vehicle (AV) planning requires a delicate balance between the efficiency of rule-based systems and the robustness of learning-based models. This paper introduces SAH-Drive (Scenario-Aware Hybrid Drive), a trajectory planner that mimics human cognitive processes by switching between "fast" rule-based logic for routine driving and "slow" diffusion-based generative modeling for complex, long-tail scenarios. Central to this approach is a biologically inspired dual-timescale decision neuron that manages the transition between planners based on environmental complexity. Evaluated on the nuPlan benchmark, SAH-Drive achieves state-of-the-art performance in challenging closed-loop environments while maintaining real-time computational feasibility.

Introduction

Modern autonomous driving systems typically rely on rule-based planners, which offer high interpretability and safety in predictable environments. However, these systems are often "brittle," failing to generalize when encountering rare or complex "long-tail" events, such as chaotic construction zones or unpredictable pedestrian behavior. Conversely, learning-based planners, particularly those utilizing diffusion models, demonstrate impressive generalization but are computationally expensive and can lack the common-sense safety constraints of hard-coded rules.

SAH-Drive addresses this dichotomy by implementing a hybrid architecture inspired by Daniel Kahneman’s "Thinking, Fast and Slow" framework. By treating rule-based logic as an intuitive, fast system and diffusion models as a deliberative, slow system, SAH-Drive optimizes for both safety and agility.

Methodology

1. Hybrid Architecture

The system integrates two distinct planning components: a state-of-the-art rule-based planner (PDM-Closed) and a Diffusion Proposal Generator. While PDM-Closed handles standard lane-following and routine maneuvers, the diffusion model acts as a creative engine, generating diverse trajectory candidates that can navigate scenarios where traditional rules might fail.

2. Dual-Timescale Decision Neuron

The core innovation of SAH-Drive is its management system, modeled after human cognitive plasticity. The Dual-Timescale Decision Neuron determines which planner should take control using two mechanisms:

  • Short-Term Plasticity: An immediate, score-based switching mechanism that reacts to sudden environmental changes.
  • Long-Term Plasticity: Based on Spike-Timing Dependent Plasticity (STDP), this mechanism reflects on the historical performance of each planner over time, allowing the system to "learn" which scenarios favor specific planning strategies.

3. Efficiency and Fusion

To ensure the system remains viable for real-time deployment, SAH-Drive utilizes a Proposal Number Regulator. In routine driving, the system reduces the number of expensive diffusion samples, relying instead on the rule-based planner. When a complex scenario is detected, the regulator increases the sample count. Finally, a Trajectory Fusion module blends the outputs of both planners to produce a final path that is both strategically sound and physically feasible.

Experimental Evaluation

Datasets and Benchmarks

The researchers utilized the nuPlan benchmark for closed-loop evaluation. The testing suite included:

  • nuPlan Mini: For initial training.
  • Val14 & Test14-Random: For routine driving assessment.
  • interPlan & Test14-Hard: Stress tests focusing on eight long-tail categories, including high-density lane changes, jaywalkers, and accident areas.

Performance Results

SAH-Drive demonstrated significant improvements over pure rule-based and pure learning-based baselines (such as PlanTF and Pluto):

  • General Performance: On the interPlan benchmark, SAH-Drive achieved a score of 64, outperforming PDM-Closed (42) and Diffusion-ES (57).
  • Long-Tail Scenarios: The system showed remarkable resilience in difficult environments. In construction zones, it scored 72 (compared to 18 for PDM-Closed). In overtaking maneuvers, it reached 78 (compared to 9 for PDM-Closed).
  • Computational Efficiency: The system maintained an average computation time of approximately 0.1s per frame, significantly faster than pure diffusion planners that often require several seconds to generate a single trajectory.

Discussion and Insights

The results highlight the importance of scenario-awareness. In standard driving conditions, the learning-based planner was active only ~6.7% of the time. However, in long-tail scenarios, its activation increased to over 43%. This shift proves that the decision neuron effectively identifies when rule-based logic is insufficient.

Furthermore, the hybrid approach mitigates the risks associated with both extremes: it avoids the over-conservatism of rules (which can lead to vehicles getting "stuck") and the over-radicalism of learning models (which can lead to collisions in high-stakes environments).

Conclusion and Future Work

SAH-Drive represents a significant step toward more human-like, adaptive autonomous planning. While the current model requires an LQR controller to ensure trajectories adhere to physical dynamics, future iterations aim to explore generative models that produce inherently constrained trajectories. Additionally, the authors plan to investigate the system's performance in even more complex, multi-agent interactive environments to further refine the "hand-over" logic between rule-based and learning-based systems.

Summary updated

BibTeX
@inproceedings{fan2025sah,
  author = {Yuqi Fan and Zhiyong Cui and Zhenning Li and Yilong Ren and Haiyang Yu},
  title = {{SAH-Drive: A Scenario-Aware Hybrid Planner for Closed-Loop Vehicle Trajectory Generation}},
  booktitle = {International Conference on Machine Learning},
  year = {2025},
  url = {https://openreview.net/forum?id=OYbZWmNHwn},
}
Figure