Abstract: Systems and methods for measuring passenger comfort and trading off passenger comfort with other metrics in autonomous vehicle planning and control. Perception of kinematic comfort relative to acceleration can vary depending on the frequency of acceleration. Systems and methods are provided to use the relationship between kinematic comfort and acceleration to introduce a cost on the ratio between acceleration and jerk. In various examples, the resulting cost can be traded off against other cost terms such as obstacle buffers and reference tracking constraints. Autonomous vehicle planning and control problem formulations can be encoded to include a direct tradeoff between both lateral and longitudinal comfort metrics and other variables.

Abstract: Autonomous vehicles may have different specialized primary planners in the vehicle planning stack that each can generate a plan for the vehicle control stack. For instance, one primary planner may be used for the vehicle to drive forward, and another primary planner may be used for the vehicle to drive in reverse. Additionally, a fallback planner may be implemented to allow the autonomous vehicle to drive in a degraded state. A unified interface is provided between the planning stack and the control stack to manage communications between the different planners in the planning stack and the control stack.

Abstract: Systems and techniques are provided for adjusting the planned path of an autonomous vehicle to converge with the actual path. An example method can include determining a first pose error associated with a first pose component from one or more pose components, wherein the first pose error is based on a first difference between an actual pose of an autonomous vehicle and a planned pose of the autonomous vehicle; determining that the first pose error associated with the first pose component exceeds a pose error threshold corresponding to the first pose component; and generating a new planned pose of the autonomous vehicle that yields a reduced first pose error associated with the first pose component, wherein the reduced first pose error is based on a second difference between the actual pose of the autonomous vehicle and the new planned pose of the autonomous vehicle.

Abstract: Disclosed are systems and techniques for managing an autonomous vehicle (AV). In some aspects, an AV may receive an engage unsupervised autonomous driving mode instruction from a remote fleet server directing the AV to engage an unsupervised autonomous driving mode. The AV may switch into the unsupervised autonomous driving mode after successful completion of one or more safety checks. In some examples, the AV may transmit a message indicating the successful completion of the one or more safety checks to the remote fleet server. The AV may receive an initiate unsupervised autonomous driving instruction from the remote fleet server directing the AV to mobilize using the unsupervised autonomous driving mode.