Research map / Probabilistic and scientific modeling / Molecules and proteins
BAT: Behavior-Aware Human-Like Trajectory Prediction for 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 BAT, a behavior-aware trajectory prediction model using dynamic geometric graphs with centrality measures, polar-coordinate pooling, and behavior/interaction/priority/position modules, evaluated on NGSIM, HighD, RounD and a new MoCAD dataset.
Problem
Predicting surrounding vehicle trajectories for autonomous driving in interactive scenarios (highways, roundabouts, urban streets) without manual driver-behavior labeling or fixed time windows.
Approach
Models driver behavior continuously via centrality measures (degree, closeness, eigenvector) on a dynamic geometric graph and their time derivatives (behavior likelihood and intensity), removing discrete behavior labels, and pools agent positions in polar coordinates instead of Cartesian grids.
Key results
- On NGSIM at 5s horizon, BAT achieves RMSE 3.62 m vs STDAN 3.67 m (overall test set).
- On HighD at 5s, BAT reaches 0.62 m vs iNATran 1.10 m and STDAN 1.66 m.
- Even with 25% of training data, BAT (25%) outperforms most baselines on NGSIM, HighD, RounD and MoCAD.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use BAT instead of STDAN | Trajectory prediction · NGSIM overall test, 5 s | RMSE (m): 3.67 -> 3.62 | Helped |
| Use BAT instead of iNATran | Trajectory prediction · HighD overall test, 5 s | RMSE (m): 1.10 -> 0.62 | Helped |
| Use BAT instead of STDAN | Trajectory prediction · RounD overall test, 4 s | RMSE (m): 2.92 -> 2.46 | Helped |
| Use BAT instead of STDAN | Trajectory prediction · MoCAD overall test, 5 s | RMSE (m): 3.32 -> 2.88 | Helped |
| Add polar-coordinate representation instead of Cartesian coordinates (ablation) | BAT · NGSIM overall test, 5 s | RMSE (m): 4.02 -> 3.62 | Helped |
| Add the behavior-aware module (ablation) | BAT · NGSIM overall test, 5 s | RMSE (m): 4.08 -> 3.62 | Helped |
| Add the interaction-aware module (ablation) | BAT · NGSIM overall test, 5 s | RMSE (m): 3.97 -> 3.62 | Helped |
| Add the priority-aware module (ablation) | BAT · NGSIM overall test, 5 s | RMSE (m): 3.97 -> 3.62 | Helped |
| Add polar-coordinate representation instead of Cartesian coordinates (ablation) | BAT · RounD overall test, 4 s | RMSE (m): 3.02 -> 2.46 | Helped |
| Add the behavior-aware module (ablation) | BAT · RounD overall test, 4 s | RMSE (m): 2.82 -> 2.46 | Helped |
| Add the interaction-aware module (ablation) | BAT · RounD overall test, 4 s | RMSE (m): 2.76 -> 2.46 | Helped |
| Add the priority-aware module (ablation) | BAT · RounD overall test, 4 s | RMSE (m): 2.68 -> 2.46 | Helped |
Surprising findings
- Trained on only 25% of the training data, BAT still beats most baselines, suggesting reduced data needs.
Limitations
- Trained on only a single hardware setup and evaluated only with RMSE and NLL; no formal efficiency (runtime) metrics reported.
- Interaction threshold distance (25 feet) chosen heuristically as average headway.
Experimental setup
Datasets: NGSIM HighD RounD MoCAD
Benchmarks: NGSIM HighD RounD MoCAD
Metrics: RMSE
Baselines: S-LSTM S-GAN CS-LSTM MHA-LSTM NLS-LSTM WSiP CF-LSTM STDAN
Builds on: multi-policy decision-making social pooling attention mechanism
Closest papers by meaning
- DriveGPT: Scaling Autoregressive Behavior Models for Driving (ICML 2025 · similarity 0.88)
- NEST: A Neuromodulated Small-world Hypergraph Trajectory Prediction Model for Autonomous Driving (AAAI 2025 · similarity 0.88)
- Prioritizing Perception-Guided Self-Supervision: A New Paradigm for Causal Modeling in End-to-End Autonomous Driving (NeurIPS 2025 · similarity 0.87)
- BehaviorGPT: Smart Agent Simulation for Autonomous Driving with Next-Patch Prediction (NeurIPS 2024 · similarity 0.87)
- UniMotion: A Unified Motion Framework for Simulation, Prediction and Planning (NeurIPS 2025 · similarity 0.87)
- Beyond Self-Play and Scale: A Behavior Benchmark for Generalization in Autonomous Driving (NeurIPS 2026 · similarity 0.86)
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.