Hierarchical Path Planning Combining DRL and Rule-Based Methods

In off-road environments, intelligent vehicles are required to execute diverse driving tasks to navigate complex and uncertain scenarios. Conventional rule-based approaches face limitations in addressing intricate situations, as they can only handle specific or simplistic cases effectively. While data-driven methods offer enhanced flexibility, they exhibit strong dependence on large-scale datasets and demonstrate inadequate generalization capabilities when encountering unfamiliar scenarios. Although reinforcement learning (RL) has paved the way for innovative paradigms in autonomous driving research, it encounters challenges such as difficult convergence and failure to generate effective decision-making strategies with the increasing complexity of driving tasks. To solve these problems, we propose a mixture decision-making framework, named A^* mix Decision Transformer (ADT), which combines Hybrid A∗ and Decision Transformer. The proposed framework integrates Hybrid A with deep reinforcement learning to improve generalization while reducing reliance on large-scale training data.