Physics-informed neural model simulates BEV parking dynamics from just 16 tests

A team of researchers (Pan, Tian, Song) has developed a physics-informed neural state-space model that accurately simulates the parking dynamics of a production battery-electric sedan. Trained on only 16 field-test maneuvers, the model captures critical low-speed behaviors that traditional kinematic idealizations ignore: actuator lag, drivetrain creep, brake-hold transitions through standstill, and frequent direction reversals. To stay physically consistent, the model imposes a gear-conditioned velocity constraint during training and learns the yaw rate as a residual on a kinematic-bicycle prior—forcing the neural network to focus on deviations from physics rather than reproducing known relationships. The approach makes conventional inference-time state limiters unnecessary. Drive, brake, and steering actuators are each handled by dedicated submodels, and the team discovered that tuning the brake model on velocity error rather than signal fidelity actually improves closed-loop performance.

The assembled command-to-vehicle chain earns Good ratings under the ISO/TS 18571 objective metric, despite being identified from minimal data. When embedded as the real-time plant in an interactive simulator, the model enables a production-representative planning stack to park the vehicle through the learned dynamics. This capability allows automotive engineers to pre-calibrate automated parking planning and control stacks entirely in virtual development, without needing the manufacturer’s proprietary chassis and actuator parameters. The code and trained models are publicly available, making the approach accessible for research and commercial validation. The work highlights how targeted physics-informed learning can dramatically reduce data requirements while maintaining high fidelity for critical low-speed maneuvers.