HyPlan: Hybrid Learning-Assisted Planning Under Uncertainty for Safe Autonomous Driving
Donald Pfaffmann, Matthias Klusch, Marcel Steinmetz
TL;DR
The paper tackles collision-free navigation under partial observability in autonomous driving by formulating the problem as a POMDP and introducing HyPlan, a hybrid learning-assisted online planner that combines multi-agent trajectory prediction, ego-path planning, a PPO-based NavPPO critic, and an IS-DESPOT* planner with confidence-based vertical pruning. The NavPPO network provides an experience-based upper bound and confidence estimates to steer planning, while CRUDE calibration aligns predictions with deployment realities. Training integrates a PPO-based objective with generalized advantage estimates to emulate planner values, and a calibration phase refines uncertainty estimates for real-time decision making. Evaluation on the CARLA-CTS2 benchmark shows HyPlan achieves safer navigation than baselines and significantly faster planning than explicit/hybrid POMDP planners, marking a meaningful step toward safer and more efficient autonomous driving under uncertainty, though it does not yet close the speed gap with pure deep learning methods.
Abstract
We present a novel hybrid learning-assisted planning method, named HyPlan, for solving the collision-free navigation problem for self-driving cars in partially observable traffic environments. HyPlan combines methods for multi-agent behavior prediction, deep reinforcement learning with proximal policy optimization and approximated online POMDP planning with heuristic confidence-based vertical pruning to reduce its execution time without compromising safety of driving. Our experimental performance analysis on the CARLA-CTS2 benchmark of critical traffic scenarios with pedestrians revealed that HyPlan may navigate safer than selected relevant baselines and perform significantly faster than considered alternative online POMDP planners.