Prediction-Driven Motion Planning: Route Integration Strategies in Attention-Based Prediction Models

TL;DR

The paper addresses conditioning multi-agent trajectory prediction on an ego route/goal to support goal-directed motion planning. It extends a transformer-based predictor (SceneMotion) with navigation integration strategies (early and late fusion) and a navigation loss, evaluated on the nuPlan dataset. Results show navigation information can improve prediction metrics and open-loop planning, with early fusion without a navigation loss providing the best trade-off between performance and complexity. The work highlights the value of navigation-aware prediction for prediction-driven planning and points to future benchmarks and command-based navigation representations.

Abstract

Combining motion prediction and motion planning offers a promising framework for enhancing interactions between automated vehicles and other traffic participants. However, this introduces challenges in conditioning predictions on navigation goals and ensuring stable, kinematically feasible trajectories. Addressing the former challenge, this paper investigates the extension of attention-based motion prediction models with navigation information. By integrating the ego vehicle's intended route and goal pose into the model architecture, we bridge the gap between multi-agent motion prediction and goal-based motion planning. We propose and evaluate several architectural navigation integration strategies to our model on the nuPlan dataset. Our results demonstrate the potential of prediction-driven motion planning, highlighting how navigation information can enhance both prediction and planning tasks. Our implementation is at: https://github.com/KIT-MRT/future-motion.

Figures (3)

• Figure 1: Motivation. The figures show three possible trajectories of the ego vehicle (red). Each trajectory is assigned a confidence, indicating how likely it is to be the actual trajectory. \ref{['fig:motivation-a']} does not include navigation information (i.e. goal and route), while \ref{['fig:motivation-b']} does.
• Figure 2: SceneMotion prediction model wagner2024scenemotion and our extensions to fuse navigation information into the model (early fusion / late fusion). The inputs highlighted in red are additional information compared to the baseline model.
• Figure 3: Left turn intersection scenario (nuPlan scenario token bec10cb7c4985dbe): A comparison of the SceneMotion baseline (top row) and the SceneMotion-A1 model (bottom row) at different times. The ground truth driven trajectory is shown in orange, while the planned trajectory are shown in blue. The ego vehicle is shown in white, while other agents are shown in green. Lanes visualized in blue indicate that they belong to the ego vehicle's route.