One of the many challenges in building robust and reliable autonomous systems is the large number of parameters and settings such systems often entail. The traditional approach to this task is simply to have system experts hand tune various parameter settings, and then validate them through simulation, offline playback, and field testing. However, this approach is tedious and time consuming for the expert, and typically produces subpar performance that does not generalize. Machine learning offers a solution to this problem in the form of learning from demonstration. Rather than ask an expert to explicitly encode his own preferences, he must simply demonstrate them, allowing the system to autonomously configure itself accordingly. This work extends this approach to the task of learning driving styles and maneuver preferences for an autonomous vehicle. Head to head experiments in simulation and with a live autonomous system demonstrate that this approach produces better autonomous performance, and with less expert interaction, than traditional hand tuning.
展开▼
