Everything is Connected (especially Planning)
Autonomous vehicle (AV) software stacks often form a feed-forward pipeline of components: perception to prediction and then planning. Such architectures help to “divide and conquer” the difficult problem of autonomous driving, where each component can be developed and debugged mostly in isolation. We argue that while such architectures are practical, they are not ideal. Ideally, we would prefer to model the AV problem as a POMDP, where perception, prediction, and planning are all inherently coupled. In this talk, we investigate some of the benefits in making small steps towards re-coupling the three components, and especially how the planning component can improve both perception and prediction, rather than just being “downstream” of them.
Bio
Rowan McAllister recently joined Toyota Research Institute as a machine learning scientist, and was previously a postdoc at the University of California, Berkeley, working with Prof. Sergey Levine on autonomous vehicle planning and reinforcement learning. He previously studied motion planning at the Australian center for Field Robotics and Bayesian modelling for data-efficient control at the University of Cambridge with Prof. Carl Rasmussen. Rowan’s contributions to the academic community include co-organizing AV workshops at ECCV, RSS, ICML, and NeurIPS.