Not-at-Fault Driving in Traffic: A Reachability-Based Approach
S. Vaskov · Hannah Larson · Shreyas Kousik · M. Johnson-Roberson · Ram Vasudevan
TL;DR
RTD-Interval computes reachable sets over time intervals, removing obstacle buffering so vehicles plan faster and less conservatively while avoiding at-fault collisions.
BRIEF
Self-driving cars must plan long maneuvers, such as an unprotected left turn across oncoming traffic, and be able to show they are not at fault if a collision occurs. An earlier method, RTD-D, precomputes the vehicle's reachable motions and plans online, but it must inflate obstacles to cover gaps between discrete time steps, which makes it conservative. The new method, RTD-Interval, computes reachable sets over whole time intervals, removing that buffer and allowing coarser time steps. In simulation it plans faster and less conservatively than RTD-D, with no at-fault collisions.
Abstract
To operate in traffic, autonomous vehicles must plan long trajectories (e.g., unprotected left turns across traffic) and validate that they are not-at-fault in a collision. Reachability-based Trajectory Design for Dynamic environments (RTD-D) is a method that plans validated trajectories, which are then tracked a feedback controller—in this work linear MPC. RTD-D computes a Forward Reachable Set (FRS) of the ego vehicle’s motion offline, then uses the FRS online to plan by discretizing time and buffering obstacles (i.e., artificially increasing their size) to compensate for the discretization; this requires choosing a conservative buffer so that the discretization is coarse enough for real-time planning. This paper eliminates the discretization-dependent buffer with a new method of computing the FRS over a prespecified choice of time intervals, allowing for a much coarser time discretization that reduces both computational cost and conservatism at runtime. The new method, RTD-Interval (RTD-I), is shown in simulation on a vehicle described by a nonlinear bicycle model in comparison to RTD-D and a State Lattice planner in unstructured dynamic environments. RTD-I is also compared to RTD-D in unprotected left turns across busy intersections, and we demonstrate that the MPC controller tracking the road centerline is unsafe on its own. RTD-I plans faster and less conservatively than RTD-D, and causes no at-fault collisions.