ADM: Accelerated Diffusion Model via Estimated Priors for Robust Motion Prediction under Uncertainties

TL;DR

ADM tackles real-time, multi-agent motion prediction under uncertainty by introducing a motion pattern estimator that learns a coarse prior to skip many diffusion denoising steps. The method combines a scenario-aware encoder, a prior-focused motion pattern estimator, and a context-conditioned denoising module to produce multiple trajectory modalities with reduced latency (≈136 ms) on Argoverse 1. A two-stage training regime using NLL and cross-entropy losses, plus targeted ablations, demonstrates that the learned prior preserves prediction quality while accelerating inference and improving robustness to perceptual noise. The approach offers a practical pathway to deploying diffusion-based trajectory forecasting in real-time autonomous driving systems and under perceptual uncertainties.

Abstract

Motion prediction is a challenging problem in autonomous driving as it demands the system to comprehend stochastic dynamics and the multi-modal nature of real-world agent interactions. Diffusion models have recently risen to prominence, and have proven particularly effective in pedestrian motion prediction tasks. However, the significant time consumption and sensitivity to noise have limited the real-time predictive capability of diffusion models. In response to these impediments, we propose a novel diffusion-based, acceleratable framework that adeptly predicts future trajectories of agents with enhanced resistance to noise. The core idea of our model is to learn a coarse-grained prior distribution of trajectory, which can skip a large number of denoise steps. This advancement not only boosts sampling efficiency but also maintains the fidelity of prediction accuracy. Our method meets the rigorous real-time operational standards essential for autonomous vehicles, enabling prompt trajectory generation that is vital for secure and efficient navigation. Through extensive experiments, our method speeds up the inference time to 136ms compared to standard diffusion model, and achieves significant improvement in multi-agent motion prediction on the Argoverse 1 motion forecasting dataset.

Figures (4)

• Figure 1: An accelerated diffusion framework featuring a motion pattern estimator for an efficient and effective trajectory denoising process.
• Figure 2: Overview of the pipeline: 1. Scenario Encoder captures interactions among agents and between agents and the map, as well as spatial-temporal relationships in each time frame to gain local embedding and global embedding. 2. Motion Pattern Estimator reparameterizes the prior distribution by learning the mean, variance, and navigation nodes. 3. Conditional Diffusion Denoising Module refines the prior distribution into a clear trajectory. 4. Probability Predictor estimates the probability for each mode of an agent's behavior. 5. Scale Net learns the scale of the Laplace distribution for regression loss.
• Figure 3: Qualitative evaluation of the ADM. The prediction samples show high accuracy in multi-agent motion prediction on the Argoverse 1 motion forecasting dataset.
• Figure 4: The effect on the model's prediction performance under different noise levels (standard deviations).