Research map / Image generation / Autoregressive and latent generation
Bidirectional Motion Transformer for Safety-Critical Traffic Scenario Generation
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.
Adv-BMT inserts a new adversarial agent with sampled collision states into real driving logs and uses a bidirectional motion transformer (BMT) to reverse-predict its history, producing diverse collision scenarios without collision training data.
Problem
Safety-critical collision scenarios are scarce in real-world driving datasets, and prior scenario generation methods yield low-diversity, low-collision-rate adversarial behaviors for testing and training autonomous driving agents.
Approach
A bidirectional motion transformer with temporally reversible motion tokens (acceleration and yaw-rate bins) predicts both future and past trajectories, so a new adversary can be placed at a sampled collision state and its history reverse-predicted, followed by rule-based rejection filtering.
Key results
- Closed-loop Adv-BMT-trained RL agents achieve 38% lower episode cost than the strongest closed-loop baseline (CAT) in the Adv-BMT environments.
- Adv-BMT attains 1.00 attack success rate versus 0.38-0.55 for baselines (CAT, SEAL, STRIVE) with higher FDD (1.51) and ADD (0.63).
- Open-loop training on Adv-BMT augmented data reduces collision rate by 14% on safety-critical environments relative to the best baseline, and refined variant reduces episode cost by 17%.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Combine reverse prediction with forward prediction | BMT · WOMD closed-loop prediction | SFDE_avg: 3.52 -> 3.29 | Helped |
| Combine reverse prediction with forward prediction | BMT · WOMD closed-loop prediction | ADD: 4.40 -> 7.00 | Helped |
| Add adversarial initialization (ablation) | BMT reverse prediction · WOMD adversarial scenario generation | FDD: 8.35 -> 9.89 | Helped |
| Add adversarial initialization (ablation) | BMT reverse prediction · WOMD adversarial scenario generation | SFDE_avg: 3.92 -> 7.17 | Hurt |
| Add rule-based filtering after adversarial initialization (ablation) | BMT reverse prediction · WOMD adversarial scenario generation | SFDE_avg: 7.17 -> 6.93 | Helped |
| Add rule-based filtering after adversarial initialization (ablation) | BMT reverse prediction · WOMD adversarial scenario generation | FDD: 9.89 -> 10.27 | Helped |
| Train TD3 on Adv-BMT-augmented scenarios rather than SEAL-augmented scenarios | TD3 · Waymo validation environments | Collision: 0.12 { $\pm$ 0.02} -> 0.11 { $\pm$ 0.02} | Helped |
| Train TD3 on Adv-BMT-augmented scenarios rather than SEAL-augmented scenarios | TD3 · Adv-BMT validation environments | Collision: 0.31 { $\pm$ 0.01} -> 0.25 { $\pm$ 0.05} | Helped |
| Add forward refinement to Adv-BMT training scenarios (ablation) | TD3 · Waymo validation environments | Reward: 31.47 { $\pm$ 3.21} -> 33.22 { $\pm$ 1.83} | Helped |
| Train with adaptive Adv-BMT rather than adaptive CAT | TD3 · Waymo validation environments | Cost Sum: 2.02 { $\pm$ 0.24} -> 1.25 { $\pm$ 0.52} | Helped |
Surprising findings
- Reverse-only prediction yields lower realism than forward prediction on WOSAC (metametric 0.554 vs 0.753), and the authors attribute the gap to training strategy.
Limitations
- High memory and computational demands due to long token sequences representing multi-agent traffic.
- Empirical study limited to a single downstream task (RL for AD agents).
Experimental setup
Datasets: Waymo Open Motion Dataset ScenarioNet format
Benchmarks: Waymo Open Motion Dataset validation Adv-BMT validation environments Waymo Open Sim Agents Challenge 2025
Metrics: SFDE SADE FDD ADD SDD JSD velocity JSD acceleration JSD TTC attack success rate collision rate
Baselines: STRIVE CAT SEAL AdvSim SafeSim
Builds on: Trajeglish SMART MetaDrive ScenarioNet TD3
Closest papers by meaning
- Generating Traffic Scenarios via In-Context Learning to Learn Better Motion Planner (AAAI 2025 · similarity 0.85)
- DiffScene: Diffusion-Based Safety-Critical Scenario Generation for Autonomous Vehicles (AAAI 2025 · similarity 0.85)
- CoC-VLA: Delving into Adversarial Domain Transfer for Explainable Autonomous Driving via Chain-of-Causality Visual-Language-Action Model (NeurIPS 2025 · similarity 0.84)
- Generating Multimodal Driving Scenes via Next-Scene Prediction (CVPR 2025 · similarity 0.84)
- BehaviorGPT: Smart Agent Simulation for Autonomous Driving with Next-Patch Prediction (NeurIPS 2024 · similarity 0.84)
- UniMotion: A Unified Motion Framework for Simulation, Prediction and Planning (NeurIPS 2025 · similarity 0.83)
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.