Machine Vision and Applications Computer Science Applications

Adaptive path planning framework for curved highway environments

Wadood, Ehsan, Andrabi, Umer Mukhtar, Kharlamov, A. A., Voronkov, I. M.
Springer Nature · 2026-10-16 · DOI 10.1007/s00138-026-01921-x

Abstract

Autonomous vehicles face critical challenges in safe and efficient path planning across curved highways, where human error accounts for 94% of traffic accidents. This paper introduces the Adaptive Path Planning Framework, a hybrid deep inverse reinforcement learning approach that integrates maximum entropy inverse reinforcement learning with deep neural networks to learn adaptive reward functions from expert demonstrations. Our framework extends traditional Markov Decision Process models by incorporating road curvature as a continuous state feature, enabling geometry-aware planning without velocity discretization—reducing state space complexity by 40%. We create a generalized nonlinear maximum entropy inverse reinforcement learning theory with proofs of uniqueness based on Hessian positivity conditions. We also propose APP-ActorCritic, a scalable model-free algorithm that combines adversarial inverse reinforcement learning with soft actor-critic updates to speed up expert imitation tasks. Comprehensive simulations on a custom SUMO-based highway benchmark show better performance: 24.2% higher success rates in collision-free navigation compared to baseline DRL methods (89.7% vs. 72.3% for DQN), 18.4% less path deviation compared to standard inverse reinforcement learning (0.81 m vs. 1.12 m), and 27% faster computation (0.11 s/episode vs. 0.18 s). The framework works 92.1% of the time on curved segments, while DQN only works 65.2% of the time. It also cuts the mean centrifugal deviation by 67%. The framework can figure out complex, geometry-aware reward functions that take into account both aggressive and conservative driving styles. This is how these gains came about. Our research shows that it is possible to connect model-based planning with data-driven imitation learning. This makes it possible to create Level 4 AV systems that prioritize safety while still driving like a human. For practical deployment, it is still important to validate simulations and integrate them with real sensor suites.

Read at publisher →