Decision-theoretic MPC: Motion Planning with Weighted Maneuver Preferences Under Uncertainty

TL;DR

This work introduces a continuous optimization based motion planner that combines multiple maneuvers by weighting the trajectory of each maneuver according to the vehicle's preferences, which eliminates the need for committing to a single maneuver.

Abstract

Continuous optimization based motion planners require specifying a maneuver class before calculating the optimal trajectory for that class. In traffic, the intentions of other participants are often unclear, presenting multiple maneuver options for the autonomous vehicle. This uncertainty can make it difficult for the vehicle to decide on the best option. This work introduces a continuous optimization based motion planner that combines multiple maneuvers by weighting the trajectory of each maneuver according to the vehicle's preferences. In this way, the planner eliminates the need for committing to a single maneuver. To maintain safety despite this increased complexity, the planner considers uncertainties ranging from perception to prediction, while ensuring the feasibility of a chance-constrained emergency maneuver. Evaluations in both driving experiments and simulation studies show enhanced interaction capabilities and comfort levels compared to conventional planners, which consider only a single maneuver.

Figures (11)

• Figure 2: An intersection crossing scene with the ego vehicle depicted in blue. Nearest conflict points of other vehicles along the ego vehicle's path are marked on a vertical axis to the right with $s_1$ and $s_2$. Considering the red vehicle's two potential routes, the point on the route that is nearest to the ego vehicle's path marks the start of the overlap. While the overlap with the red vehicle is limited, the overlap with the green vehicle could potentially continue indefinitely, as it follows the ego vehicle's route.
• Figure 5: Uncertainty projection in conjunction with the MPCC formulation. The ego vehicle, depicted in blue, is modeled using three circles. Its position is projected onto the closest centerline segment. Lateral and longitudinal errors between the reference contouring point and the projected point are denoted with ${e}_{\text{lat}}$ and ${e}_{\text{lon}}$, respectively. The uncertainty of the other vehicle's position is modeled by an ellipse and is projected onto the closest centerline segment.
• Figure 6: Trajectories at an intersection crossing for various sensor ranges.
• Figure 7: Executed speed profiles between the decision-theoretic MPC ( ) and a conventional MPC ( ) in the phantom object detection scenario. While the conventional planner calculates the motion by assuming the phantom to be valid, the decision-theoretic planner assesses the likelihood of a phantom presence.
• Figure 8: Safe intersection crossing in the presence of rule-violating participants. The autonomous gray vehicle has the right of way. However, the speed of the oncoming black vehicle suggests it will not yield. The autonomous vehicle reduces its speed to satisfy the yield maneuver before the intersection. Once it becomes clear that the oncoming vehicle will yield, it accelerates back to its desired driving speed. Frames are incremented every $0.25s$.

Theorems & Definitions (4)

• Lemma 1
• proof
• Lemma 2
• proof