Safe Receding Horizon Path Planning for Autonomous Vehicles * Tom Schouwenaars † Eric Feron † Jonathan How § † Laboratory for Information and Decision Systems § Space Systems Laboratory Massachusetts Institute of Technology Cambridge, MA 02139 {toms, feron, jhow}@mit.edu Abstract This paper extends a recently developed approach to optimal path planning of autonomous vehicles to account for safety. Recent research in the field of trajectory optimization for Unmanned Aerial Vehicles (UAV’s) is based on the use of mixed integer linear programming (MILP) to account for obstacle and collision avoidance constraints. To allow for real-time computation, a receding horizon approach was proposed, in which a path is computed online by solving a MILP over a limited horizon at each time step. However, satisfying anti-collision constraints over this limited planning horizon at a certain instant in time does not necessarily yield a collision free path during future time steps. In this paper, a receding horizon algorithm is presented that ensures a priori safety for a single vehicle with lim- ited knowledge of the environment, if only stationary obstacles are present. More precisely, it guarantees that the vehicle can always transition to a safe mode, by maintaining a feasible path to a predefined safe state at each time step. 1 Introduction In recent years, the space and defense community has gained considerable interest in fully autonomous space- and aircraft systems, or so called Unmanned Aerial Vehicles (UAV’s). Such systems need no or minor human control from a ground station, thereby reducing operating costs and enabling missions in harsh or remote environments. A significant portion of the autonomy consists of path planning capabilities: the problem is to guide the vehicle through an obstacle field, taking into account its dynamic and kinematic properties. It has been proven that the motion planning problem is intrinsically NP-hard [1]. Namely, the space of possible control actions is extremely large, requiring simplifica- tions to reduce the dimensionality, when a solution needs to be found in practical time. In recent years, several methods based on randomized algorithms, such as Probabilistic * Research funded under the DARPA Software Enabled Control Program, AFRL contract # F33615-01-C-1850. Tom Schouwenaars was partly supported by a Francqui Fellowship of the Belgian American Educational Foundation.