Research map / Probabilistic and scientific modeling / Molecules and proteins
CaDeT: a Causal Disentanglement Approach for Robust Trajectory Prediction in Autonomous Driving
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes CaDeT, a causal disentanglement trajectory predictor on a dynamic heterogeneous graph that uses uncertainty-driven latent interventions and a variance-based invariance loss to improve robustness to spurious correlations.
Problem
Learning-based trajectory prediction for autonomous driving overfits to spurious correlations and is brittle under small perturbations and distribution shifts.
Approach
Disentangle causal and spurious spatiotemporal patterns with attention masks over a DyHIN scene graph, simulate distribution shifts by intervening on spurious features using uncertain feature statistics sampled from mini-batch variance, and penalize the variance of prediction loss across interventions to focus on causal patterns.
Key results
- On AGV2 leaderboard, CaDeT reaches minFDE6 1.24 and MR6 0.15, state-of-the-art on most metrics versus QCNet 1.29 and 0.16.
- Under adversarial perturbations on AGV2 at 100% training data, CaDeT's PRS is 87.59% (targeted) vs 78.51% for Forecast-MAE, and maintains about 84.58% at 50% data where Forecast-MAE falls to 65.96%.
- On WOMD causal-perturbation evaluation, CaDeT achieves 12% higher PRS than the best baseline MTR.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add causal disentanglement to the DyHIN baseline (ablation) | DyHIN trajectory prediction · AGV2 validation | minADE6: 0.93 -> 0.88 | Helped |
| add same-timestamp spatial intervention with causal disentanglement (ablation) | DyHIN trajectory prediction · AGV2 validation | minADE6: 0.93 -> 0.77 | Helped |
| add same-agent temporal intervention with causal disentanglement (ablation) | DyHIN trajectory prediction · AGV2 validation | minADE6: 0.93 -> 0.75 | Helped |
| use uncertainty-driven intervention instead of random pattern replacement (ablation) | CaDeT · AGV2 validation | minADE6: 0.80 -> 0.71 | Helped |
| use CaDeT instead of QCNet† | AGV2 motion forecasting · AGV2 leaderboard | minFDE6: 1.19 -> 1.24 | Helped |
| use CaDeT instead of Forecast-MAE | trajectory prediction under adversarial perturbations · AGV2 validation, d = 100%, targeted | PRS: 78.51 -> 87.59 | Helped |
| use CaDeT instead of Forecast-MAE | trajectory prediction in an unseen domain · AGV2 unseen test domain | minADE6: 0.897 -> 0.638 | Helped |
| use CaDeT instead of MTR | trajectory prediction after removing non-causal agents · WOMD validation, Trained-All | PRS: 80 (%) -> 92 (%) | Helped |
Surprising findings
- Ensembled QCNet still beats CaDeT on several AGV2 leaderboard metrics, so CaDeT is not uniformly best on accuracy.
Limitations
- Accuracy on AGV2 is not uniformly best; the 1.86 b-minFDE6 trails the QCNet ensemble and the MR1 results trail some ensembles.
- Robustness evaluation relies on a specific adversarial attack setup and a causal-agent annotation subset, so generality is uncertain.
- Hyperparameter lambda is sensitive and dataset-specific (1 for AGV2, 0.1 for WOMD).
Experimental setup
Datasets: Argoverse-2 motion forecasting Waymo Open Motion Dataset
Benchmarks: Argoverse-2 motion forecasting leaderboard Waymo Open Motion Dataset causal annotations
Metrics: minADE6 minFDE6 b-minFDE6 MR6 minADE1 minFDE1 MR1 mAP perturbation resistance score (PRS)
Baselines: THOMAS MTR GANet GoRela FRM QCNet ProphNet Forecast-MAE HPTR HDGT Multipath++ SceneTransformer
Builds on: DyHGT invariant risk minimization style variance penalty PointNet Forecast-MAE
Closest papers by meaning
- BAT: Behavior-Aware Human-Like Trajectory Prediction for Autonomous Driving (AAAI 2024 · similarity 0.86)
- LaKD: Length-agnostic Knowledge Distillation for Trajectory Prediction with Any Length Observations (NeurIPS 2024 · similarity 0.85)
- Prioritizing Perception-Guided Self-Supervision: A New Paradigm for Causal Modeling in End-to-End Autonomous Driving (NeurIPS 2025 · similarity 0.85)
- C2F-TP: A Coarse-to-Fine Denoising Framework for Uncertainty-Aware Trajectory Prediction (AAAI 2025 · similarity 0.85)
- Certified Human Trajectory Prediction (CVPR 2025 · similarity 0.85)
- Learning Cooperative Trajectory Representations for Motion Forecasting (NeurIPS 2024 · similarity 0.85)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.