Multi-Task Trajectory Prediction Using a Vehicle-Lane Disentangled Conditional Variational Autoencoder.

Prior research has shown that integrating dynamic vehicle states with static high-definition (HD) maps is essential for modeling complex agent-scene interactions in autonomous driving. Conventional approaches typically rely on static scene encodings that fail to capture the evolving nature of spatial contexts during motion. These existing frameworks often utilize unstructured latent spaces, which restricts their capacity to generate diverse yet contextually coherent future paths. The inability to align map features with vehicle movement leads to suboptimal scene semantics. Such limitations hinder the effectiveness of multi-modal prediction systems in dense urban environments. The absence of structured latent models makes it difficult to separate the specific intentions of an agent from the physical constraints imposed by the road layout. This absence of evidence motivated a new approach to handle time-aware scene alignment and structured latent modeling.

MS-SLV introduces a generative framework designed to enhance trajectory forecasting through a vehicle-lane disentangled Conditional Variational Autoencoder (CVAE). The system aims to align high-definition (HD) map features with vehicle motion to capture evolving scene semantics accurately. Researchers sought to create a structured latent model that explicitly separates agent-specific intent from scene-level constraints. This architecture incorporates an auxiliary lane prediction task to provide targeted supervision for improved scene understanding. By jointly predicting future trajectories and lane sequences, the framework enables more interpretable and scene-consistent forecasts. The study focuses on overcoming the limitations of unstructured latent variables in generative models for autonomous navigation. Engineers designed this multi-task approach to ensure that predicted paths remain within the boundaries of the high-definition (HD) map.

The investigative process utilized the MS-SLV generative framework to process multimodal information for autonomous navigation. Engineers implemented a time-aware scene encoder to synchronize high-definition (HD) map features with the dynamic states of the vehicle. The experimental design featured a structured latent model to achieve disentanglement between agent intent and environmental constraints. To refine latent variable learning, the team integrated an auxiliary lane prediction task into the training pipeline. Evaluation of the proposed architecture was conducted using the comprehensive nuScenes dataset, a standard benchmark for autonomous driving research. This methodology allowed for the joint estimation of future trajectories and corresponding lane sequences. The researchers compared the performance of this vehicle-lane disentangled Conditional Variational Autoencoder (CVAE) against several state-of-the-art baselines.

MS-SLV achieved a 12.37% reduction in average displacement error (ADE) compared to existing state-of-the-art methods. The framework also demonstrated a 7.67% improvement in final displacement error (FDE) during extensive evaluations on the nuScenes dataset. Multi-modal prediction performance saw significant gains, with the top-5 Miss Rate (MR5) decreasing by 26%. The top-10 Miss Rate (MR10) showed an even greater reduction of 33% relative to the strongest baseline evaluated. The system successfully lowered the Off-Road Rate (ORR) by 3%, indicating superior scene consistency. These metrics confirm that the structured latent model effectively captures diverse and contextually appropriate future paths. The results highlight the effectiveness of aligning high-definition (HD) map features with temporal vehicle motion.

The findings suggest that disentangling agent intent from scene constraints significantly enhances the reliability of trajectory forecasting in autonomous vehicles. This approach provides a pathway for developing more interpretable and scene-consistent navigation systems. The integration of auxiliary tasks like lane prediction offers a robust method for supervising scene understanding in generative models. Future research may build upon this time-aware scene encoding to handle even more complex agent-scene interactions. These advancements contribute to the safety and efficiency of self-driving technology by reducing displacement errors and miss rates. The study establishes MS-SLV as a high-performing framework for multi-task trajectory prediction. By addressing the gap in structured latent modeling, this work provides a foundation for more sophisticated agent-scene interaction models.