Kinematics-Constrained Reinforcement Learning for Adversarial Driving Scenario Generation

Abstract

Learning-based adversarial strategies can discover collision-inducing maneuvers, yet the generated behaviors become physically implausible when low-level execution relies on trajectory overwrites that bypass vehicle dynamics. This disconnect between the intended maneuver and its execution breaks kinematic consistency among displacement, velocity, and acceleration, producing teleportation artifacts and boundviolating accelerations. To ensure kinematic consistency, we propose a reinforcement learning framework based on Implicit Q-Learning (IQL). Within this framework, the learned policy produces smooth, physically bounded acceleration commands, while a constraint layer forbids any coordinate overwrite to enforce continuous motion. To encourage physically valid control, the policy is trained with a stabilized time-to-collision shaping term and normalized acceleration and jerk penalties. To prevent kinematic anomalies from contaminating the training data, we introduce a Kinematic Anomaly Detector (KAD) that filters out transitions with acceleration bound violations or discrepancies between displacement-derived and simulator-reported velocities. Experiments on the TeraSim platform demonstrate that our approach substantially improves kinematic validity while maintaining comparable adversarial efficiency, reducing the kinematic invalid rate from 29.5% to 13.0%.