Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Adaptive Prediction Ensemble: Improving Out-of-Distribution Generalization of Motion Forecasting

Jinning Li, Jiachen Li, Sangjae Bae, David Isele

TL;DR

This work tackles zero-shot generalization in autonomous driving motion forecasting, where deep predictors falter on out-of-distribution data. It introduces Adaptive Prediction Ensemble (APE), a three-component framework combining a deep learning predictor, a rule-based baseline, and a learned routing function trained concurrently to select the most reliable output. Empirical results on Waymo Open Motion Dataset and Argoverse show that APE improves cross-dataset generalization, especially for long-horizon predictions, outperforming individual predictors and simple routing baselines. The findings suggest that hybrid, routing-based strategies can yield robust, dataset-agnostic motion predictions with practical implications for safer autonomous navigation.

Abstract

Deep learning-based trajectory prediction models for autonomous driving often struggle with generalization to out-of-distribution (OOD) scenarios, sometimes performing worse than simple rule-based models. To address this limitation, we propose a novel framework, Adaptive Prediction Ensemble (APE), which integrates deep learning and rule-based prediction experts. A learned routing function, trained concurrently with the deep learning model, dynamically selects the most reliable prediction based on the input scenario. Our experiments on large-scale datasets, including Waymo Open Motion Dataset (WOMD) and Argoverse, demonstrate improvement in zero-shot generalization across datasets. We show that our method outperforms individual prediction models and other variants, particularly in long-horizon prediction and scenarios with a high proportion of OOD data. This work highlights the potential of hybrid approaches for robust and generalizable motion prediction in autonomous driving. More details can be found on the project page: https://sites.google.com/view/ape-generalization.

Adaptive Prediction Ensemble: Improving Out-of-Distribution Generalization of Motion Forecasting

TL;DR

This work tackles zero-shot generalization in autonomous driving motion forecasting, where deep predictors falter on out-of-distribution data. It introduces Adaptive Prediction Ensemble (APE), a three-component framework combining a deep learning predictor, a rule-based baseline, and a learned routing function trained concurrently to select the most reliable output. Empirical results on Waymo Open Motion Dataset and Argoverse show that APE improves cross-dataset generalization, especially for long-horizon predictions, outperforming individual predictors and simple routing baselines. The findings suggest that hybrid, routing-based strategies can yield robust, dataset-agnostic motion predictions with practical implications for safer autonomous navigation.

Abstract

Deep learning-based trajectory prediction models for autonomous driving often struggle with generalization to out-of-distribution (OOD) scenarios, sometimes performing worse than simple rule-based models. To address this limitation, we propose a novel framework, Adaptive Prediction Ensemble (APE), which integrates deep learning and rule-based prediction experts. A learned routing function, trained concurrently with the deep learning model, dynamically selects the most reliable prediction based on the input scenario. Our experiments on large-scale datasets, including Waymo Open Motion Dataset (WOMD) and Argoverse, demonstrate improvement in zero-shot generalization across datasets. We show that our method outperforms individual prediction models and other variants, particularly in long-horizon prediction and scenarios with a high proportion of OOD data. This work highlights the potential of hybrid approaches for robust and generalizable motion prediction in autonomous driving. More details can be found on the project page: https://sites.google.com/view/ape-generalization.
Paper Structure (25 sections, 3 equations, 4 figures, 1 table, 1 algorithm)

This paper contains 25 sections, 3 equations, 4 figures, 1 table, 1 algorithm.

Table of Contents

  1. Introduction
  2. Related Work
  3. Improving Out-of-Distribution Motion Prediction
  4. Mixture-of-Experts
  5. Finetuning with Human Feedback
  6. Problem Formulation
  7. Adaptive Prediction Ensemble
  8. Deep Learning Prediction Expert
  9. Rule-Based Prediction Expert
  10. Learned Routing Function
  11. Practical Implementation
  12. Experiments
  13. Experiment Setting
  14. Deep Learning Prediction Expert
  15. Rule-Based Prediction Expert
  16. ...and 10 more sections

Figures (4)

  • Figure 1: Illustration of the motivation of improving prediction algorithm by Adaptive Prediction Ensemble. (a) An example scenario where vanilla rule-based prediction algorithm outperforms deep NN prediction algorithm (MTR shi2022motion). (b) A comparison of the error (minADE) between deep NN and rule-based prediction. The rule-based method outperforms deep NN in a considerable amount of scenarios, which are the ones below the red line.
  • Figure 2: The model structure of the learned routing function and the deep learning-based prediction algorithm, which share the same backbone of scene encoder, and are trained concurrently. In this way, the routing function shares the same level of powerful scene understanding ability with the motion prediction algorithm, while trained concurrently on all footprint prediction outputs increases its exposure to diverse anomalous trajectory predictions and hence more capability on differentiating prediction quality.
  • Figure 3: The trajectory prediction visualization curated by the learning-based routing function. (a)(b) Cases where MTR generalizes better than the constant velocity model. (c)(d) Cases where the constant velocity model generalizes better than MTR.
  • Figure 4: The performance gain percentage vs. (a) OOD data percentage in the test dataset; and (b) Prediction horizon. The performance gain is monotonically increasing in both cases, indicating that our method has more advantage over individual predictors when OOD data is common in the test dataset and the task horizon is long.