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

TL; DR

Adv-BMT for data augmentation: augments real-world driving logs with realistic and diverse collision interactions.

Adv-BMT is a two-stage generation pipeline: adversarial initialization + reverse motion prediction to produce plausible collision scenarios.

Adv-BMT is an adversarial scenario generator for closed-loop learning: continuously produces diverse, realistic, collision-seeking interactions that target the ego vehicle.

Adv-BMT

Adv-BMT is a two-staged pipeline: first, it initializes diverse collision states between a new adversary agent and ego vehicle; then, it reconstructs the adversarial trajectories via BMT’s reverse predictions. A rule-based rejection sampling mechanism is used to filter candidate trajectories from unsatisfactory adversarial initializations, and maintain realistic collision interactions. In the output, the new agent maintains realistic interactions with surrounding traffic.

BMT Architecture

BMT is a transformer-based motion prediction model, and is able to predict both future and history trajectories for vehicles, bicycles, and pedestrain agents.

Diverse Adversarial Behaviors

Safer Agent via Adversarial Learnings

Adversarial RL Evaluations in Waymo Environments

Adversarial RL Evaluated in Safety-critical Environments

Prior Works

CAT (CoRL 2022): Closed-loop adversarial training for safe end-to-end driving.

MetaDrive (TPAMI 2021): An open-source platform for large-scale traffic scenario simulation and modeling