Bilevel Inverse Optimal Control for Lane-Change Costs in Automated Driving

Manual tuning of trajectory planners for advanced driver assistance systems is time-consuming and difficult to scale. Learning cost functions from demonstrations offers a viable data-driven alternative, but real-world lane-change data is affected by noisy measurements and missing reliable temporal segmentation methods. This paper studies the learning of quadratic cost functions for a linear parameter-varying kinematic model in lane-change scenarios from real-world driving data. We adapt an inverse optimal control-based segmentation approach to detect the lane-change start and use it to calibrate a heuristic segmentation rule. For cost function learning, we use bilevel inverse optimal control to accommodate incomplete state observations with statistical consistency under measurement noise and apply it to a large dataset using stochastic gradient descent. We also learn a parameter-varying cost function with the parameters represented by a neural network. The results show that both the curvature reference and the number of optimized entries in the cost matrices affect reconstruction, and that parameter-varying cost functions further improve reconstruction accuracy.