AI for Autonomous Robotic Cars

•• perception and sensor fusion, mapping and localization, modeling and learning, and prediction and decision making. My current research touches on several of these topics, with the goal of developing algorithms that let robotic vehicles sense and understand the environment and drive safely in accordance with human rules and conventions. Recently, one of my main interests has been path planning for autonomous navigation in unknown environments. The task here is to generate trajectories in real time in response to new information about the environment obtained by the robot’s onboard sensors. A main challenge in developing a practical path planner arises from the fact that the space of all vehicle controls (and thus trajectories) is continuous. So, the task of computing near-optimal trajectories that are safe and smooth and that satisfy the vehicle’s kinematic constraints becomes a complex nonlinear continuous-variable optimization problem. My colleagues from the Stanford Racing Team (http://cs.stanford.edu/group/ roadrunner) and I have developed a pathplanning algorithm that addresses this computational problem by combining heuristic graph search, potential fi elds, and numerical continuous-variable optimization. We’ve successfully tested this research in the DARPA Urban Challenge (www.darpa. mil/grandchallenge), in which our vehicle flperformed maneuvers in freenavigation environments. Another research direction that’s of great interest to me and, in fact, a prerequisite for good decision making is the problem of sensing and understanding the environment. A main challenge in this area is the sparsity and noise in the sensor data, which lead to an inherently ambiguous problem and force the use of probabilistic inference techniques for perception and mapping. Along this direction, my colleagues and I recently have been investigating ways to infer global properties of the environment on the basis of local observations provided by the sensors. For example, many human-built environments (such as parking lots) are well structured. So, I’ve been working on methods for constructing a globally consistent model of that structure by extracting local geometric features from sensor data and then using probabilistic techniques to combine them into a global view. Despite the tremendous breakthroughs that have been made in autonomous driving in recent years, many unsolved problems remain. Among those that especially interest me are reliable identifi cation and tracking of dynamic objects, predicting their motion, and using that information when making decisions. Making progress on these tasks will require advances on several fronts, including robust perception techniques, methods for learning and modeling dynamical systems, and effi cient planning and decision-making algorithms. My goal over the next several years is to continue tackling these problems, using probabilistic AI techniques.