Augmenting Safety-Critical Driving Scenarios while Preserving Similarity to Expert Trajectories

TL;DR

This work tackles distributional shift in imitation learning for safety-critical driving by focusing on minority, safety-critical trajectories. It introduces a cluster-based trajectory augmentation framework that preserves similarity to expert data, using an LSTM-based autoencoder to embed trajectories and K-means to form clusters, followed by synthesizing new trajectories within clusters via a geometric transform that preserves start/end points and guide-shape, along with stringent quality checks. The approach is validated on urban (InD) and highway (TrafficJams) datasets, showing improved closed-loop performance and robustness over baselines such as SMOTE-upsampled augmentation. Key contributions include identifying safety-critical clusters through AE embeddings, a principled within-cluster trajectory synthesis method, and comprehensive safety-focused QA checks that ensure feasibility of augmented data. The results demonstrate practical impact for data-efficient, safer autonomous driving planning under distributional shift.

Abstract

Trajectory augmentation serves as a means to mitigate distributional shift in imitation learning. However, imitating trajectories that inadequately represent the original expert data can result in undesirable behaviors, particularly in safety-critical scenarios. We propose a trajectory augmentation method designed to maintain similarity with expert trajectory data. To accomplish this, we first cluster trajectories to identify minority yet safety-critical groups. Then, we combine the trajectories within the same cluster through geometrical transformation to create new trajectories. These trajectories are then added to the training dataset, provided that they meet our specified safety-related criteria. Our experiments exhibit that training an imitation learning model using these augmented trajectories can significantly improve closed-loop performance.

Figures (6)

• Figure 1: Synthesis steps for generating a new trajectory using our defined transformative process applied to two trajectories within each pair: (a) one satisfying predefined acceptance criteria for augmentation, and (b) one failing to meet the specified acceptance criteria for augmentation.
• Figure 2: Architecture of our LSTM-based autoencoder.
• Figure 3: t-SNE of the embeddings and their corresponding clusters.
• Figure 4: Examples of the results of our trajectory clustering approach. (a) Sample clustered headings, representing approaching intersection for turn left (left), leaving intersection after turn left (center), and keeping lane (right). (b) Examples of clustered velocities, representing acceleration (left), deceleration (center), and monotonous velocity (right).
• Figure 5: Comparison of data cluster percentages before and after augmentation using our approach, for our urban driving scenarios (top) and highway driving scenarios (bottom).