Planning with Adaptive World Models for Autonomous Driving

arXiv 2024

Carnegie Mellon University

tl;dr: We propose a model-predictive-control (MPC)-based planner (AdaptiveDriver) that makes use of world models who's agent behaviours adapt to the sequence. We compare AdaptiveDriver to state-of-art baselines (PDMC[1]).

PDMC collides with a decelerating lead vehicle (frame 20) while AdaptiveDriver avoids the collision.

PDMC collides with a decelerating lead vehicle (frame 20) while AdaptiveDriver avoids the collision.

AdaptiveDriver reaches the goal state (green square) faster than PDMC for a turn-right sequence.

AdaptiveDriver reaches the goal state (green square) faster than PDMC for a straight-road sequence.

More Complex Scenarios (Lane change, Unprotected Cross turns) and Failure Case.

Lane Change scenario: AdaptiveDriver changes lane to the right smoothly and goes towards the goal state (green square) compared to PDMC for a change-lane sequence.

Unprotected Cross Turn scenario: PDMC collides with a vehicle when turning left on an unprotected cross turn (frame 45) while AdaptiveDriver avoids the collision.

Unprotected Right Turn scenario: AdaptiveDriver reaches the goal state (green square) faster than PDMC for a turn-right sequence.

Traversing Intersection scenario: PDMC collides with a decelerating lead vehicle (frame 94) while AdaptiveDriver avoids the collision.

Failure Cases.

Traversing Intersection scenario: PDMC and Adaptive Driver collides with a vehicle at cross roads (frame 45). Both are moving at high speeds and decelerating to late-detected lead vehicle made it hard and this may have lead to collision.

Lane Change scenario: PDMC and Adaptive Driver fail to change the lane though they reach the destination on the parallel lane safely.

Related Works

[1] Dauner, Daniel, et al. "Parting with misconceptions about learning-based vehicle motion planning." Conference on Robot Learning. PMLR, 2023.