Research map / Probabilistic and scientific modeling / Molecules and proteins
CCTR: Calibrating Trajectory Prediction for Uncertainty-Aware Motion Planning 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 CCTR, a plug-and-play framework that regularizes trajectory predictors' variance toward squared prediction error and learns per-prediction temperature scalers from context and history, improving calibration and motion planning on Argoverse.
Problem
Trajectory predictors used in autonomous driving produce over-confident or unreliable variance estimates, and existing calibration methods use a single global temperature, which hurts downstream motion planning.
Approach
Combine a calibration-based regularizer that aligns predicted variance with squared prediction error during training with a post-processing network that outputs customized per-prediction temperature ratios from kinematic, social, and spatiotemporal (BEV CNN+GRU) features, trained on the validation set with rescaled NLL.
Key results
- With ED-LSTM, CCTR reduces ECE by 15% vs. original predictor and 11% vs. ETS (0.1335 vs 0.1567 and 0.1493).
- With ED-LSTM and RRT* planner, CCTR decreases l2 planning error by 46% vs. original and 17% vs. ETS (4.6017 vs 8.6258 original).
- CCTR keeps or slightly improves prediction accuracy (e.g., ED-LSTM ADE 1.74 to 1.74, FDE 3.99 to 3.95).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use CCTR calibration instead of ETS | ED-LSTM · Argoverse | ECE: 0.1493 -> 0.1335 | Helped |
| Use CCTR calibration instead of ETS | ED-LSTM with RRT* planning · Argoverse | ℓ2 error (RRT*): 5.5453 -> 4.6017 | Helped |
| Use CCTR-EN rather than E2E planning | ED-LSTM + NMP · Argoverse | ℓ2 error: 2.3012 -> 2.1459 | Helped |
| Add the kinematics feature extractor (ablation) | CCTR with ED-LSTM · Argoverse | ECE: 0.1395 -> 0.1335 | Helped |
| Add the social feature extractor (ablation) | CCTR with ED-LSTM · Argoverse | ECE: 0.1402 -> 0.1335 | Helped |
| Add the spatiotemporal feature extractor (ablation) | CCTR with ED-LSTM · Argoverse | ECE: 0.1420 -> 0.1335 | Helped |
| Add the calibration regularizer (ablation) | CCTR with ED-LSTM · Argoverse | ECE: 0.1367 -> 0.1335 | Helped |
| Add the post-processing module (ablation) | CCTR with ED-LSTM · Argoverse | ECE: 0.1468 -> 0.1335 | Helped |
Surprising findings
- About 14.37% of predictions are under-confident even though most predictions are over-confident, so a single global enlarging scaler is insufficient.
- Original predictors show essentially no correlation between variance and ADE (-0.01), whereas CCTR yields a 0.95 correlation.
Limitations
- Calibration model is trained on the validation set, requiring held-out data.
- Only evaluated on Argoverse with five predictors and two planners; the paper states future work could use more advanced post-processing modules.
Experimental setup
Datasets: Argoverse
Benchmarks: Argoverse
Metrics: ECE MCE NCE l2 error ADE FDE
Baselines: Temperature Scaling Isotonic Regression Ensemble Temperature Scaling MCDropout Deep Ensemble E2E FF
Builds on: temperature scaling ensemble temperature scaling isotonic regression calibration NLL training for distributional regression
Closest papers by meaning
- C2F-TP: A Coarse-to-Fine Denoising Framework for Uncertainty-Aware Trajectory Prediction (AAAI 2025 · similarity 0.86)
- CUQDS: Conformal Uncertainty Quantification Under Distribution Shift for Trajectory Prediction (AAAI 2025 · similarity 0.84)
- Evaluation of Trajectory Distribution Predictions with Energy Score (ICML 2024 · similarity 0.84)
- Don't Shake the Wheel: Momentum-Aware Planning in End-to-End Autonomous Driving (CVPR 2025 · similarity 0.84)
- Certified Human Trajectory Prediction (CVPR 2025 · similarity 0.84)
- Producing and Leveraging Online Map Uncertainty in Trajectory Prediction (CVPR 2024 · similarity 0.84)
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.