Adv-BMT: Bidirectional Motion Transformer for Safety-Critical Traffic Scenario Generation

2025-06-11

AI Summary

Safety-critical long-tail scenarios are scarce in autonomous driving testing, and real-world data often fails to cover adversarial collision conditions. Method: This paper proposes the Bidirectional Motion Transformer (BMT) framework—a novel inverse temporal motion modeling paradigm that requires no collision-labeled pretraining. BMT enables end-to-end, unsupervised generation of safety-critical scenarios via adversarial initialization and implicit dynamics learning grounded in proxy state representations. Contribution/Results: BMT overcomes conventional reliance on annotated data or simulation priors, significantly enhancing scenario diversity and physical plausibility. Experiments show that augmenting training with BMT-generated data reduces per-trip collision rates by 20% for autonomous driving models, effectively alleviating the long-tail data bottleneck and improving scenario-based verification efficacy.

Technology Category

• Multiagent Systems: Adversarial Agents
• Machine Learning: Adversarial Learning & Robustness
• Search and Optimization: Adversarial Search

Application Category

• Responsible Web: Machine-in-the-loop, human agency and autonomy
• User Modeling, Personalization and Recommendation: Attacks and countermeasures in recommendation systems
• Semantics and Knowledge: Data modeling to support human-machine intelligence, including LLMs agents, intelligent system behavior, explanations, and user-friendly interactions

Abstract

Scenario-based testing is essential for validating the performance of autonomous driving (AD) systems. However, such testing is limited by the scarcity of long-tailed, safety-critical scenarios in existing datasets collected in the real world. To tackle the data issue, we propose the Adv-BMT framework, which augments real-world scenarios with diverse and realistic adversarial interactions. The core component of Adv-BMT is a bidirectional motion transformer (BMT) model to perform inverse traffic motion predictions, which takes agent information in the last time step of the scenario as input, and reconstruct the traffic in the inverse of chronological order until the initial time step. The Adv-BMT framework is a two-staged pipeline: it first conducts adversarial initializations and then inverse motion predictions. Different from previous work, we do not need any collision data for pretraining, and are able to generate realistic and diverse collision interactions. Our experimental results validate the quality of generated collision scenarios by Adv-BMT: training in our augmented dataset would reduce episode collision rates by 20% compared to previous work.

Problem

• Generates safety-critical traffic scenarios for autonomous driving testing
• Augments real-world datasets with diverse adversarial interactions
• Improves collision scenario quality without pretraining collision data

Innovation

• Bidirectional motion transformer for inverse predictions
• Two-staged adversarial and inverse prediction pipeline
• Generates realistic collisions without pretraining data