Learning maneuver dictionaries for ground robot planning

Vehicle dynamics is typically handled by models whose parameters are found through system identification or manually computed from the vehicle's characteristics. While these methods provide accurate theoretical dynamical models, they may not take into account differences between individual vehicles, lack adaptability to new environments and may not handle sophisticated models, requiring hand-crafted heuristics for backwards motion for example. Similarly to space and aerial maneuver-based planning methods, we demonstrate a simple and computationally fast planning method for ground robots with obstacle avoidance. It bypasses the need for model parameters identification and hand-crafted heuristics, learns the particularities of individual vehicles, allows on-line adaptation and sophisticated models. Human-driven or autonomouslydriven trajectories are recorded and stored into a trajectory bank. While in learning mode, the robot records each traveled trajectory and places it into a bank, indexed by the initial speeds of each left and right wheels and the ending position at a fixed radius. Only the best trajectories are stored in the trajectory bank and then reused during autonomous runs for optimal short-range planning. Pre-computed (but not recorded) trajectories have been used in previous work and provide an important computational advantage over on-line computation methods, which are less practical in real-time applications due to the high-dimensional search space. A collision-free platform was developed without any hand-crafted heuristics or knowledge about the vehicle's characteristics. This method is demonstrated on the LAGR platform, a non-holonomic (differential drive) off-road mobile robot.