Real Time Continuous Curvature Path Planner for an Autonomous Vehicle in an Urban Environment

An real-time algorithm based on Ariadne’s Clew and A* is developed to plan dynamically feasible trajectories appropriate to a vehicle traversing an urban environment in a legal manner. Simple precomputed clothoids are used to build a tree of possible maneuvers which is searched to minimize a cost function depending on path length, steering effort and lane discipline. The algorithm dynamically adapts its map sampling resolution to suit the difficulty of the problem. Complex situations where reversing is required are handled naturally. A proof-of-concept prototype real-time velocity planning algorithm is demonstrated. This functionality is proposed and recommended for use as part of Team Caltech’s entry into the DARPA Urban Challenge 2007. In November 2007 DARPA will hold the Urban Challenge race, in which qualifying autonomous vehicles will attempt to navigate a route through an emulated urban environment to specified checkpoints in a safe, legal manner. It is the aim of this project to develop Team Caltech’s existing software onboard the 2005 DARPA Grand Challenge (DGC) entry, Alice, to be able to plan a dynamically feasible, traffic-law abidding trajectory. The existing path planning module is designed for a static, desert environment. The new planner will have to deal with moving obstacles whilst also changing its emphasis from minimising traversal time to performing safe, legal driving. This report describes the development of a spatial planning algorithm that could be used either to produce appropriate seeds for the optimization stage of the existing framework or as part of a newly proposed spatial-temporal planner framework. The planning module developed for the Grand Challenge 2005 has two stages: stage 1 samples a speed limit map over a grid aligned to the Route Data Definition File (RDDF) corridor, and performs a forward-only graph search to find the fastest route; stage 2 employs the Non-Linear Programming optimization package SNOPT to minimize traversal time in a receding horizon framework. More details can be found in [1]. This project focuses on development of the stage 1 planner. Various planning methodologies have been considered and two candidates are proposed. The first is an receding horizon time traversal optimization framework, based on the 2005 DGC software but with dynamic obstacle avoidance as incorporated by Martin Larsson [2]. Various problems are anticipated with this approach: time traversal minimisation is difficult to reconcile with safe, predictable driving how to incorporate legal constraints the receding distance horizon framework is does not allow for many urban situations such as queuing no method exists to add costs to actions such as crossing lanes the gradient descent method struggles with the binary obstacles found in urban environments unnecessarily complicated for highly constrained urban situations There are also problems generating a useful seed for optimization. The rationale behind the existing split into two stages is that the stage 1 planner approximates a solution considering only spatial constraints. However, to generate a useful seed when dynamic obstacles are involved we must consider the coupling between the spatial and temporal problems: whether we need to change direction to avoid another vehicle depends upon the time we get there, which in turn depends on the spatial route we took. To generate the seed the full spatialtemporal problem must be solved, so the complexity is not reduced. Larsson [2]