CAFE-AD: Cross-Scenario Adaptive Feature Enhancement for Trajectory Planning in Autonomous Driving

Junrui Zhang, Chenjie Wang, Jie Peng, Haoyu Li, Jianmin Ji, Yu Zhang, Yanyong Zhang

Introduction

Imitation learning (IL) has emerged as a dominant paradigm for developing autonomous driving planners, leveraging large-scale human driving datasets like nuPlan to mimic expert behavior. Despite its success, IL-based planners often struggle in closed-loop testing due to two primary phenomena: causal confusion and the long-tail distribution of driving scenarios. Causal confusion occurs when a model incorrectly associates irrelevant environmental cues with the ego-vehicle's actions. Simultaneously, the long-tail distribution—where simple, stationary scenarios dominate the training data—often results in "lazy" planners that fail to execute proactive maneuvers in complex situations.

To address these challenges, we present CAFE-AD (Cross-scenario Adaptive Feature Enhancement), a framework designed to enhance feature representation through targeted noise reduction and scenario balancing.

Methodology

CAFE-AD builds upon a Transformer-based architecture (using PLUTO as a baseline), introducing two core modules: the Adaptive Pruning Transformer (APT) Encoder and the Cross-Scenario Feature Interpolation (CSFI) module.

Adaptive Pruning Transformer (APT) Encoder

The APT-Encoder is designed to mitigate causal confusion by filtering out non-causal environmental noise. In dense urban environments, a planner is often overwhelmed by tokens representing distant vehicles or irrelevant map features. The APT-Encoder employs a token-ranking mechanism based on attention probabilities from the ego-vehicle's perspective. By identifying and physically pruning tokens and attention connections deemed unimportant, the model is forced to focus on the critical causal factors—such as a vehicle merging into the ego-path—thereby reducing the risk of overfitting to environmental noise.

Cross-Scenario Feature Interpolation (CSFI)

The CSFI module addresses the long-tail distribution problem by performing data augmentation in the feature space. Standard datasets are often skewed; for instance, nearly 45% of the nuPlan dataset consists of stationary or near-stationary scenarios. This imbalance leads models to favor safe but inactive behaviors.

CSFI utilizes a scenario classifier to decompose encoded features into two components:

  1. Scenario-relevant features: Characteristics specific to a maneuver (e.g., the trajectory curvature of a left turn).
  2. Scenario-generic features: General driving attributes applicable across various contexts.

By interpolating scenario-relevant features from dominant classes with those from rarer, high-stakes scenarios, the module generates synthetic diversity during training. This prevents the model from simply memorizing the "stationary" behavior of the majority and encourages the learning of robust maneuvers for rare edge cases.

Experimental Evaluation

The performance of CAFE-AD was evaluated on the nuPlan dataset, specifically focusing on the nuPlan Test14-Hard benchmark. This benchmark is composed of the 20 most challenging scenarios across 14 distinct categories, such as unprotected left turns and complex lane changes, providing a rigorous test for closed-loop planning.

Key Results

CAFE-AD demonstrated state-of-the-art (SOTA) performance across several key metrics:

  • SOTA Performance: CAFE-AD achieved an NR-Score of 76.04, significantly outperforming the previous leader, PLUTO (73.87), as well as rule-based (PDM-Closed) and pure-learning (PlanTF) planners.
  • Proactivity and Progress: A critical finding was the improvement in the Progress metric. CAFE-AD reached a score of 77.51, compared to PLUTO’s 75.07. This indicates that the model is less prone to the "lazy" behavior typical of IL models, successfully navigating through traffic rather than remaining stationary to avoid potential collisions.
  • Ablation Studies: Experiments removing the APT-Encoder and CSFI modules confirmed that both noise reduction and feature-space interpolation are vital for the observed performance gains.

Discussion and Conclusion

The success of CAFE-AD highlights the importance of treating attention mechanisms as more than just weighting tools; they can serve as active filters to combat causal confusion. Furthermore, the decomposition of features into generic and specific components allows for a more nuanced approach to data augmentation than traditional global methods.

While CAFE-AD shows significant promise, its performance is currently linked to the accuracy of its scenario classifier and the tuning of the pruning ratio ($\pi_p$). Future work will explore the integration of Large Language Models (LLMs) to provide more sophisticated semantic scene understanding, potentially replacing discrete scenario classification with continuous, descriptive reasoning. In conclusion, CAFE-AD provides a robust framework for improving the safety and proactivity of autonomous planners, as evidenced by its SOTA results on challenging benchmarks and successful validation on physical autonomous platforms.

Summary updated

BibTeX
@inproceedings{zhang2025cafe,
  author = {Junrui Zhang and Chenjie Wang and Jie Peng and Haoyu Li and Jianmin Ji and Yu Zhang and Yanyong Zhang},
  title = {{CAFE-AD: Cross-Scenario Adaptive Feature Enhancement for Trajectory Planning in Autonomous Driving}},
  booktitle = {IEEE International Conference on Robotics and Automation},
  year = {2025},
  doi = {10.1109/icra55743.2025.11127589},
}
Figure