A Generalizable Physics-Guided Causal Model for Trajectory Prediction in Autonomous Driving
Zhenyu Zong, Yuchen Wang, Haohong Lin, Lu Gan, Huajie Shao
AI summary
Problem
Data-driven trajectory prediction models fail to generalize to unseen domains due to spurious correlations and city-specific environmental features. Existing domain generalization techniques are computationally expensive, data-hungry, or lack physical plausibility for zero-shot scenarios.
Approach
The method disentangles domain-invariant scene representations from map data using causal intervention, then fuses them with a two-wheel kinematic model via a causal attention decoder to generate physically plausible trajectories.
Key results
- Proposes PCM with a Disentangled Scene Encoder and CausalODE Decoder
- Extracts domain-invariant features without causal annotations or predefined structures
- Achieves superior zero-shot generalization across nuPlan and WOMD datasets
- Significantly outperforms baselines on minADE, minFDE, and miss rate metrics
Why it matters
Enables autonomous vehicles to safely predict agent movements in completely new urban environments without costly retraining or fine-tuning.
Abstract
Trajectory prediction for traffic agents is critical for safe autonomous driving. However, achieving effective zero- shot generalization in previously unseen domains remains a significant challenge. Motivated by the consistent nature of kinematics across diverse domains, we aim to incorporate domain-invariant knowledge to enhance zero-shot trajectory prediction capabilities. The key challenges include: 1) effec- tively extracting domain-invariant scene representations, and 2) integrating invariant features with kinematic models to enable generalized predictions. To address these challenges, we propose a novel generalizable Physics-guided Causal Model (PCM), which comprises two core components: a Disentangled Scene Encoder, which adopts intervention-based disentangle- ment to extract domain-invariant features from scenes, and a CausalODE Decoder, which employs a causal attention mecha- nism to effectively integrate kinematic models with meaningful contextual information. Extensive experiments on real-world autonomous driving datasets demonstrate our method’s su- perior zero-shot generalization performance in unseen cities, significantly outperforming competitive baselines. The source code is released at https://github.com/ZY-Zong/Phys ics-guided-Causal-Model.