![[2602.21319] Uncertainty-Aware Diffusion Model for Multimodal Highway Trajectory Prediction via DDIM Sampling](https://arxiv.org/static/browse/0.3.4/images/arxiv-logo-fb.png){kind=logo}
[2602.21319] Uncertainty-Aware Diffusion Model for Multimodal Highway Trajectory Prediction via DDIM Sampling
Summary
The paper presents cVMDx, an advanced diffusion model for multimodal highway trajectory prediction, enhancing efficiency and accuracy in autonomous driving scenarios.
Why It Matters
As autonomous driving technology evolves, accurate trajectory prediction is crucial for safety and efficiency. This study addresses existing limitations in trajectory prediction models, offering a solution that significantly reduces inference time while improving predictive capabilities, which is vital for real-world applications.
Key Takeaways
- cVMDx improves trajectory prediction efficiency by up to 100x.
- The model enhances robustness and multimodal predictive capabilities.
- Utilizes DDIM sampling for practical multi-sample generation.
- Incorporates a Gaussian Mixture Model for tractable multimodal predictions.
- Demonstrated higher accuracy on the highD dataset compared to previous models.
Computer Science > Machine Learning arXiv:2602.21319 (cs) [Submitted on 24 Feb 2026] Title:Uncertainty-Aware Diffusion Model for Multimodal Highway Trajectory Prediction via DDIM Sampling Authors:Marion Neumeier, Niklas Roßberg, Michael Botsch, Wolfgang Utschick View a PDF of the paper titled Uncertainty-Aware Diffusion Model for Multimodal Highway Trajectory Prediction via DDIM Sampling, by Marion Neumeier and Niklas Ro{\ss}berg and Michael Botsch and Wolfgang Utschick View PDF Abstract:Accurate and uncertainty-aware trajectory prediction remains a core challenge for autonomous driving, driven by complex multi-agent interactions, diverse scene contexts and the inherently stochastic nature of future motion. Diffusion-based generative models have recently shown strong potential for capturing multimodal futures, yet existing approaches such as cVMD suffer from slow sampling, limited exploitation of generative diversity and brittle scenario encodings. This work introduces cVMDx, an enhanced diffusion-based trajectory prediction framework that improves efficiency, robustness and multimodal predictive capability. Through DDIM sampling, cVMDx achieves up to a 100x reduction in inference time, enabling practical multi-sample generation for uncertainty estimation. A fitted Gaussian Mixture Model further provides tractable multimodal predictions from the generated trajectories. In addition, a CVQ-VAE variant is evaluated for scenario encoding. Experiments on the publicly available ...