Adaptive diffusion model sharpens multi-agent trajectory prediction for autonomous driving

Researchers have unveiled a new artificial intelligence framework that predicts how multiple vehicles will move through traffic at once, using a generative technique known as a conditional diffusion model. The system, called CDJMP, short for Conditional Diffusion model-based Joint Motion Prediction, was described in a paper published on 1 September 2026 in the journal Autonomous Intelligent Systems. In benchmark tests on two widely used driving datasets, the framework achieved state-of-the-art results, reducing the minimum average displacement error by up to 9 percent and the minimum final displacement error by up to 10 percent on the INTERACTION dataset, while also cutting inference time substantially compared with fixed-step diffusion baselines.

Trajectory prediction sits at the heart of autonomous driving. A self-driving vehicle must anticipate not only where a single nearby car will go, but how every agent in a crowded scene will negotiate, yield, merge, or compete for space. The authors of the new study note that multi-agent joint motion prediction serves as a critical bridge between perception and decision-making modules in autonomous driving systems, and that existing methods have struggled to model complex agent interactions while generating reliable multi-modal trajectories. Multi-modality matters because the future is genuinely uncertain: a car at an intersection may turn left, turn right, or continue straight, and a predictor that collapses all of those possibilities into a single averaged path is of little use to a planner.

Diffusion models, the same family of generative techniques behind many modern image generators, have increasingly been adopted for trajectory prediction because they learn the joint probability distribution of future motion through an iterative denoising process. By adding noise to candidate trajectories and then learning to remove it step by step, conditioned on the observed scene, these models can maintain diversity among their predictions while capturing intricate spatio-temporal interaction patterns. Yet the authors identify two persistent problems. First, conventional diffusion models often fail to balance the diversity of generated trajectories against the accuracy of prediction, particularly when generating joint futures for many agents simultaneously. Second, diffusion models are slow: classical approaches require numerous iterative denoising steps, producing latency that is unsuitable for real-time autonomous driving applications, where systems typically require prediction latency within 100 milliseconds.

CDJMP tackles both problems with a two-stage design. In the first stage, a probabilistic initializer predicts proposal trajectories together with an estimate of how many denoising steps each proposal actually needs. Rather than starting from random Gaussian noise, the diffusion process begins from the model’s own predicted trajectory, to which Gaussian noise is added to form the noisy input. The initializer decomposes the proposal distribution into a mean, a variance, and sample predictions, all produced by trainable multilayer perceptron modules. Crucially, the variance estimate encodes the model’s own uncertainty about the motion, and that uncertainty signal drives the adaptive diffusion-step module, which dynamically determines the reverse diffusion depth for each candidate trajectory instead of applying a fixed number of steps to everything.

The second stage introduces a group-aware conditional encoder that captures dynamic multi-agent interactions and guides the diffusion process toward coherent multi-modal outcomes. The encoder defines, for each target agent, a set of dynamic surrounding agents selected according to both an overlapping time window and Euclidean distance, ensuring that only temporally co-existing neighbors within a certain influence radius are considered. In the experiments, that influence radius was set to 20 meters for both datasets. Grouping masks generated from spatial proximity control which agent pairs participate in the attention computation, and a grouped cross-attention module then produces an interaction feature for each ego-surrounding agent pair. The resulting group-aware conditional embedding is fed into the denoising network as a conditioning input, so that the refinement of every trajectory hypothesis is informed by the behavior of the agents around it.

The adaptive diffusion-step mechanism is not merely an engineering heuristic; the authors provide a probabilistic justification. They reinterpret the diffusion step as a latent random variable representing the required denoising depth conditioned on scene complexity, and model the predictive process as a marginalization over candidate steps. Because no ground-truth annotation of the correct number of denoising steps exists, the module is optimized in an unsupervised variational manner, maximizing an evidence lower bound that combines a reconstruction term with an entropy regularization term. The entropy term explicitly quantifies the model’s uncertainty regarding the appropriate denoising depth. During inference, the step with the highest posterior probability is selected from a candidate list, allowing simple scenes to be refined with fewer iterations and complex ones with more.

The experimental evaluation relied on two demanding benchmarks. The INTERACTION dataset is an adversarial and cooperative motion dataset covering interactive driving scenarios from multiple countries, containing behaviors such as negotiation and aggressive decision making, with 425,192 training and 104,627 test trajectory sequences, one second of observation history and three seconds of prediction at 10 hertz. The Argoverse Motion Forecasting dataset, jointly released by Argo AI, Carnegie Mellon University, and the Georgia Institute of Technology, offers 324,557 driving scenarios, with 205,942 training and 78,143 test sequences, two seconds of history and three seconds of future prediction. On INTERACTION, CDJMP achieved a minimum average displacement error of 0.2695, a minimum final displacement error of 0.8434, and a miss rate of 0.0695, the best performance across all evaluation metrics. Compared with the previous best-performing method, HyperMTP, CDJMP reduced the average displacement error by 9 percent, the final displacement error by 10 percent, and the miss rate by a remarkable 65 percent. On Argoverse, the model attained a minimum average displacement error of 0.6413, a minimum final displacement error of 1.0288, and a miss rate of 0.1456, outperforming all baselines, although the authors note that the relatively higher miss rate there reflects the model’s sensitivity to map-induced uncertainty, since the current version of CDJMP does not explicitly incorporate map priors.

Ablation studies isolated the contribution of each component. Starting from a diffusion-based baseline following the LED formulation, adding the group-aware conditional encoder reduced the three-second average displacement error on INTERACTION from 0.4263 to 0.3886, an improvement of roughly 9 percent, and by about 11 percent on Argoverse. Adding the adaptive diffusion-step module on top brought the three-second final displacement error on INTERACTION down by around 22 percent, from 1.0816 to 0.8434, and further reduced the three-second average displacement error on Argoverse by approximately 15 percent. The authors interpret these gains as evidence that the encoder strengthens interaction modeling while the adaptive step module refines denoising by adjusting the diffusion depth according to motion uncertainty, mitigating cumulative prediction errors over longer horizons.

The efficiency results are arguably the most striking. In a comparison of inference times on the Argoverse dataset, sampled from 1000 test samples with means and standard deviations reported, the model without the adaptive diffusion-step module showed inference time increasing significantly as the number of denoising steps grew, whereas the full model maintained the best prediction accuracy while achieving substantially lower inference latency and small standard deviations, indicating good inference stability. In absolute terms, CDJMP achieves a single-sample inference time of 29.60 milliseconds on a single NVIDIA RTX 4090 GPU, well within the 100-millisecond budget that autonomous driving systems typically require, whereas the fixed-step baseline with 30 denoising steps reaches 374.78 milliseconds, violating real-time constraints. For multi-agent scenarios with 10 agents, parallel batch inference processes all agents simultaneously via GPU parallelism, yielding approximately 31.8 milliseconds on typical INTERACTION scenes. The authors acknowledge that scalability under extremely dense scenarios with more than 50 agents remains an open challenge for future work.

Qualitative visualizations on INTERACTION scenarios, including interaction, merge, and roundabout scenes, showed the method producing more coherent trajectory patterns than the baseline and capturing diverse motion behaviors of multiple agents with strong robustness across traffic scenarios. The framework also avoids reliance on high-definition map inputs, which allowed fair comparison against map-free baselines such as LED and HyperMTP under identical observation horizons, prediction horizons, and six-mode evaluation protocols. Looking ahead, the authors plan to extend CDJMP to heterogeneous traffic scenarios with mixed human-driven and autonomous vehicles, to investigate online adaptation strategies for previously unseen environments, and to integrate high-definition map information and traffic rules into the diffusion-based framework to enhance scene understanding and trajectory plausibility. Code for the model has been released in a public repository, and the open-access paper offers full methodological detail, including the variational derivation of the adaptive step mechanism, for researchers working on the perception-to-planning pipeline of autonomous vehicles.

Subject of Research: Adaptive conditional diffusion models for multi-agent joint motion prediction in autonomous driving

Article Title: CDJMP: an adaptive conditional diffusion model for multi-agent joint motion prediction in autonomous driving

Article References: Zheng, W., Huang, C., Zhang, H., & Yin, H. (2026). CDJMP: an adaptive conditional diffusion model for multi-agent joint motion prediction in autonomous driving. Autonomous Intelligent Systems, 6(1), Article 18. https://doi.org/10.1007/s43684-026-00138-z

Image Credits: AI Generated

DOI: 10.1007/s43684-026-00138-z

Keywords: autonomous driving, trajectory prediction, diffusion models, multi-agent systems, motion forecasting, machine learning, INTERACTION dataset, Argoverse dataset, inference efficiency, variational inference, generative AI, intelligent vehicles