Abstract:

A chance-constrained probabilistic robust trajectory planning and closed-loop control method was proposed to resolve contradictions between perception uncertainty and planning feasibility in high-speed constrained driving scenarios. The study modeled perception uncertainty via a multi-dimensional Gaussian distribution, collision probability was transformed into a deterministic dynamic safety boundary based on the 3σ principle, enabling adaptive safety margins. A PyTorch-based fully differentiable non-convex solver was developed for gradient optimization of trajectories under strict dynamic and spatial constraints. The results show that, evaluated in a CARLA closed-loop framework, the proposed method overcomes the 27% collision rate of deterministic planning and the 99% failure rate of conservative strategies, achieving a 100% success rate under set conditions. Even under an extreme + 0.6 m perception error, the maximum lateral tracking error remains within 0.174 m without collisions. Ultimately, the proposed method breaks the “no-solution” deadlock in narrow roads, validating its strong engineering executability and safety.

Key words: automotive active safety, trajectory planning, perception uncertainty, chance constrained, spatially constrained scenarios, robust optimization