SafeShift: Safety-Informed Distribution Shifts for Robust Trajectory Prediction in Autonomous Driving
SafeShift uses counterfactual probing to identify safety-critical scenarios, reducing collision rates by 10% in autonomous driving trajectory prediction.
Key Findings
Methodology
SafeShift introduces a hierarchical scenario characterization framework combining multi-modal feature extraction with counterfactual future trajectory inference. It leverages scene features—both individual and social—to compute safety relevance scores. By simulating hypothetical scenarios where proactive maneuvers are absent, the method assesses latent risks. These scores facilitate the creation of safety-informed distribution shifts, enabling models to be trained and evaluated on more challenging, safety-critical scenes. Experiments on WOMD demonstrate that this approach effectively identifies risky scenarios and enhances model robustness, achieving a 10% reduction in collision rates.
Key Results
- The proposed scoring mechanism successfully detects high-risk scenes, leading to a 10% average reduction in collision rates across models trained on WOMD. The models show improved ADE, FDE, and mAP metrics when evaluated on out-of-distribution safety-critical scenes.
- In out-of-distribution tests, models trained with safety-informed shifts outperform baseline models, especially in complex interactions such as U-turns and VRU navigation, confirming the effectiveness of scene scoring.
- Counterfactual future trajectory extrapolation enhances the models’ ability to anticipate and avoid risky behaviors, significantly improving safety metrics without sacrificing overall prediction accuracy.
Significance
This work addresses a critical gap in autonomous driving safety research by systematically identifying and utilizing latent risk scenarios within real-world datasets. It moves beyond traditional data augmentation, providing a scalable, data-driven approach to improve model robustness against rare but dangerous events. The integration of counterfactual reasoning into scene scoring offers a novel pathway to preemptively recognize hazards, potentially reducing accidents and fatalities in real-world deployment. Its methodology bridges the gap between simulation-based safety testing and real-world data, offering a practical solution for industry adoption and safety assurance.
Technical Contribution
The core technical innovation lies in the multi-layered scene feature extraction combined with a counterfactual scoring scheme that assesses safety relevance. The framework incorporates a hierarchical aggregation of individual and social features, with a novel future extrapolation process in the Frenet frame to simulate unacted scenarios. This scoring guides the creation of safety-informed distribution shifts and enables model remediations through weighted loss functions and auxiliary safety-aware features. These contributions collectively advance the state-of-the-art in robust trajectory prediction and safety-critical scenario detection.
Novelty
This is the first comprehensive framework integrating counterfactual scene scoring with distribution shift construction for safety-critical scenario identification in real-world datasets. Unlike prior works focusing on static scene features or heuristic heuristics, SafeShift systematically quantifies latent risks via hypothetical future trajectories, enabling precise risk segmentation. Its approach to leveraging existing datasets without synthetic scene generation marks a significant departure from simulation-based methods, offering practical scalability and real-world relevance.
Limitations
- The accuracy of scene scoring heavily depends on the quality of feature extraction; noisy or incomplete data can impair risk detection. The assumption of constant velocity in future extrapolation may oversimplify dynamic environments, limiting effectiveness in highly unpredictable scenarios.
- The method's reliance on static scene features may overlook environmental changes, such as weather or lighting, which also influence safety. Computational costs for large-scale feature extraction and counterfactual inference could hinder real-time deployment.
- Further validation across diverse datasets and complex scenarios is needed to confirm generalizability. Integration with multi-modal sensor data and dynamic environment modeling remains future work.
Future Work
Future research will focus on integrating multi-sensor data for richer scene understanding, developing adaptive future inference models that account for environmental dynamics, and extending the framework to multi-task learning for joint safety and behavior prediction. Additionally, real-time implementation and large-scale deployment studies are planned to validate practical effectiveness.
AI Executive Summary
Autonomous driving systems face a fundamental challenge: ensuring safety in unpredictable, complex traffic environments. While large-scale datasets like WOMD provide valuable real-world scenarios, they often lack sufficient safety-critical events, limiting models’ ability to handle rare but dangerous situations. Traditional simulation-based approaches to generate such scenarios suffer from realism gaps, and real-world testing poses safety risks. To address this, SafeShift introduces a novel framework that leverages existing datasets by identifying and scoring safety-relevant scenarios through counterfactual probing.
This approach involves extracting multi-modal scene features, including individual vehicle behaviors and inter-agent interactions, then applying a hierarchical scoring system that assesses safety-criticality. By simulating hypothetical future trajectories where proactive maneuvers are absent, the method uncovers latent risks that may be overlooked in recorded data. These scores enable the construction of safety-informed distribution shifts, which are used to train and evaluate models under more challenging conditions.
Experimental results on WOMD demonstrate that models trained with safety-shifted data significantly reduce collision rates by an average of 10%. The scoring mechanism effectively isolates scenarios with high potential for accidents, especially in complex maneuvers like U-turns and VRU interactions. Incorporating these scenarios into training enhances model robustness, making autonomous systems better prepared for real-world hazards.
Despite its success, the framework faces limitations such as reliance on feature quality and simplified future extrapolation assumptions. Future work aims to incorporate multi-modal sensor data, dynamic environment modeling, and real-time deployment. Overall, SafeShift offers a scalable, data-driven pathway to improve autonomous vehicle safety, bridging the gap between dataset limitations and real-world safety requirements.
Deep Dive
Abstract
As autonomous driving technology matures, safety and robustness of its key components, including trajectory prediction, is vital. Though real-world datasets, such as Waymo Open Motion, provide realistic recorded scenarios for model development, they often lack truly safety-critical situations. Rather than utilizing unrealistic simulation or dangerous real-world testing, we instead propose a framework to characterize such datasets and find hidden safety-relevant scenarios within. Our approach expands the spectrum of safety-relevance, allowing us to study trajectory prediction models under a safety-informed, distribution shift setting. We contribute a generalized scenario characterization method, a novel scoring scheme to find subtly-avoided risky scenarios, and an evaluation of trajectory prediction models in this setting. We further contribute a remediation strategy, achieving a 10% average reduction in prediction collision rates. To facilitate future research, we release our code to the public: github.com/cmubig/SafeShift