HybridNav: Hybrid Path Planning for Autonomous Agricultural Headland Turns

Abstract

The automation of agricultural tractors is hindered by a critical algorithmic conflict between achieving optimal straight-line row coverage and executing kinematically constrained U-turns in narrow, cluttered headlands. Traditional heuristic planners generate mathematically optimal paths that violate physical steering limits, whereas pure reinforcement learning approaches lack the global spatial awareness to guarantee complete field coverage. Furthermore, existing hierarchical planners (such as A* coupled with reactive avoidance or Model Predictive Control) suffer from either local-minima entrapment or prohibitive online computational latency. To resolve these challenges, we introduce HybridNav, a principled two-layer motion planning framework designed to ensure both theoretical completeness and continuous physical drivability. The global layer utilizes the A* algorithm over a 0.25-meter resolution occupancy grid to rapidly compute complete boustrophedon coverage across 27 crop row strips (effective swath width 2.96 m with 1.2% overlap across the 80 m interior field width). A transition controller monitors vehicle proximity and hands operational command to a Soft Actor-Critic (SAC) reinforcement learning agent at an analytically derived threshold of τh = 11.0 m before the row boundary. Unlike prior blind turning models, the local SAC agent incorporates an augmented 16-dimensional state vector that explicitly encodes an 8-beam rangefinder obstacle profile alongside vehicle kinematics, enabling collision-free maneuvering around headland clutter. Trained over 250,000 simulated timesteps using curriculum learning and domain randomization, the SAC agent executes fluid U-turns within the strictly confined 10-meter headland margin. Furthermore, to overcome the limitations of discrete three-point geometric approximations, this paper introduces a continuous Ackermann Compliance Metric (ηAcont) that evaluates both instantaneous path curvature and steering slew rate along the continuous trajectory. Benchmarked against classical baselines (APF, GA, RRT*, A*) as well as alternative hierarchical paradigms (A*+DWA, Hybrid A*+MPC), HybridNav achieves 100% field coverage, zero collisions, and 98.4% continuous Ackermann compliance with sub-2 ms inference latency.