ScholarWorks@숙명여자대학교

초록

Local path planning for autonomous driving often generates multiple candidates and selects a collision-free one based on geometric clearance over a finite prediction horizon. However, such geometric screening implicitly assumes ideal execution, and the resulting decision can be inconsistent with real closed-loop behavior when tracking deviations and model–plant mismatch occur. This paper proposes a control-aware planning framework that incorporates the closed-loop tracking envelope of a tube-based MPC tracker into collision decision making. At each step, the planner generates kinematically feasible quintic-polynomial candidates from an anchor on the previously selected path and discretizes them on the same prediction grid used by the controller. To enable real-time operation, we adopt a two-stage screening architecture. Stage 1 performs fast geometric gating using a bounding-box-based collision check and forwards a reduced set of promising candidates to the next stage using a lightweight proxy cost that promotes goal compliance and temporal consistency. Stage 2 then performs control-aware verification for each remaining candidate by solving a nominal tracking MPC with the candidate as the reference, using the candidate-conditioned predicted lateral deviation to shift the ego footprint toward the expected trajectory center (pose correction), and inflating the footprint by a certified, speed-parametric tube radius derived from the robust invariant set underlying tube MPC to cover bounded residual effects. The resulting shift-and-inflate screening rejects candidates that would collide under realistic closed-loop deviations while avoiding overly conservative symmetric margins. High-fidelity CarSim/Simulink simulations with static obstacles and dynamic cut-in scenarios demonstrate that the proposed method improves closed-loop safety and robustness compared to purely geometric screening and static-margin baselines under aggressive, high-speed maneuvers.

키워드