Boundary-Guided Trajectory Prediction for Road Aware and Physically Feasible Autonomous Driving

Abstract: Accurate prediction of surrounding road users' trajectories is essential for safe and efficient autonomous driving. While deep learning models have improved performance, challenges remain in preventing off-road predictions and ensuring kinematic feasibility. Existing methods incorporate road-awareness modules and enforce kinematic constraints but lack plausibility guarantees and often introduce trade-offs in complexity and flexibility. This paper proposes a novel framework that formulates trajectory prediction as a constrained regression guided by permissible driving directions and their boundaries. Using the agent's current state and an HD map, our approach defines the valid boundaries and ensures on-road predictions by training the network to learn superimposed paths between left and right boundary polylines. To guarantee feasibility, the model predicts acceleration profiles that determine the vehicle's travel distance along these paths while adhering to kinematic constraints. We evaluate our approach on the Argoverse-2 dataset against the HPTR baseline. Our approach shows a slight decrease in benchmark metrics compared to HPTR but notably improves final displacement error and eliminates infeasible trajectories. Moreover, the proposed approach has superior generalization to less prevalent maneuvers and unseen out-of-distribution scenarios, reducing the off-road rate under adversarial attacks from 66\% to just 1\%. These results highlight the effectiveness of our approach in generating feasible and robust predictions.

Boundary-Guided Trajectory Prediction for Autonomous Driving

The paper "Boundary-Guided Trajectory Prediction for Road Aware and Physically Feasible Autonomous Driving" introduces a novel methodology for predicting the trajectories of surrounding road users in the context of autonomous driving. This approach addresses two critical challenges: ensuring road-awareness and maintaining kinematic feasibility. The authors propose a trajectory prediction framework informed by permissible driving directions and their boundaries, derived from high-definition (HD) maps and the current state of the vehicle. This method seeks to integrate both spatial awareness and adherence to physical constraints within a unified model.

Core Contributions

1. Boundary Set Generation: The authors introduce an algorithm for extracting the boundaries of permissible driving directions from HD maps. This is achieved by formulating the HD map as a directed graph, identifying start lanes based on the agent's position and heading, and employing reachability analysis to determine potential goal lanes. The algorithm generates a boundary set consisting of left and right polylines, providing a spatial frame for valid driving maneuvers.

2. Integration into Network Architecture: The proposed approach is embedded within a transformer-based architecture, adapted for this study as HPTR$_{\text{bd}}$. Boundary embeddings are incorporated into the network's polyline encoder and interaction encoder to facilitate informed trajectory predictions. By utilizing superposition weights for left and right boundary points, the network predicts a superimposed path, which ensures on-road adherence.

3. Physical Feasibility through Kinematic Constraints: A pure pursuit layer, coupled with predicted acceleration profiles, transforms superimposed paths into kinematically feasible trajectories. This layer guarantees that the computed trajectories adhere to vehicle dynamics and steering constraints, addressing a common shortcoming in conventional deep learning models.

Evaluation and Results

The authors evaluate their framework on the Argoverse-2 dataset, comparing the performance against baseline models like the original HPTR and an adapted kinematic model (HPTR$_{\text{kin}}$). While there is a slight decrease in benchmark metrics (e.g., minADE and minFDE) due to the introduction of feasibility constraints, the model shows significant improvements in final displacement error and the elimination of infeasible trajectories. In particular, the model reduces the off-road rate under adversarial attacks from 66% to a mere 1%.

Additionally, the maneuver-based analysis highlights the model's superiority in handling complex maneuvers such as U-turns, demonstrating its potential for generalizing across diverse road scenarios not present in training. This robust generalization capability is further corroborated by the model's response to adversarial perturbations, which mimic challenging road scenarios, emphasizing its enhanced road awareness.

Implications and Future Work

Practically, this paper's methodology offers significant advantages for trajectory prediction in autonomous vehicles, notably its ability to incorporate road constraints effectively and its potential to generalize across various driving contexts and road typologies. Theoretically, it introduces a paradigm of using boundary-guided pathways to enhance model interpretability and adherence to physical and spatial constraints.

For future research, expanding this constrained prediction approach to include pedestrians and cyclists, who exhibit less predictable and non-lane-bound behavior, remains an intriguing domain. Moreover, integrating this predictive model with planning and control systems in autonomous vehicles could further validate its utility and robustness in dynamic real-world environments. Additionally, enhancing the boundary set generation to account for more complex urban driving scenarios with dense road topologies would further bolster the practical applicability of the proposed framework.