A Multi-Stage Time-Variant Motion Planner for Agile Autonomous Driving Maneuvers

Agile, dynamic, and safe motion planning in interactive environments is a key challenge for autonomous vehicles. Most research has focused on enhancing established planning algorithms, often aiming to improve optimality or computational efficiency within fixed planning horizons. This article introduces a novel sampling-based planner that employs a multi-stage approach to generate a tree of maneuvers by concatenating trajectory segments of different time horizons and selecting the best maneuver. To enable real-time performance, we reduce computational complexity through several strategies such as limiting the number of planning stages and using informed sampling. This allows us to combine short agile maneuvers with longer-horizon trajectories, enhancing recursive feasibility while respecting dynamical constraints. We evaluate our algorithm in simulation using high-speed, highly agile autonomous racing scenarios, and compare it against a state-of-the-art trajectory planner. Our experiments show that the approach can perform static and dynamic obstacle avoidance and succeed in challenging overtaking scenarios. These findings indicate that our method improves the agility and maneuverability of autonomous vehicles, expanding their operational domain under challenging conditions.