自主决策能力是无人驾驶技术的核心，现阶段的无人驾驶基于规则的决策方法面临因真实交通场景状态维度高、不确定性强等特点导致的决策失效的挑战。自学习的决策方法能够从历史驾驶数据中自主学习，并适应高维、高不确定性的交通场景，这使得自学习算法具有进一步提升无人驾驶决策算法表现的潜力。然而自学习决策的决策过程难以解释，这与无人驾驶汽车的高可靠性需求相矛盾。针对以上问题，本研究提出了混合规则策略与自学习策略的无人驾驶决策方法，该方法能够在随机复杂交通场景中实现自主学习且安全可控。本研究首先提出了无人驾驶汽车规则与自学习平行混合决策方法，该方法设计了规则决策系统与自学习决策系统平行决策的方案，通过激活函数建立自学习策略和规则策略的联系，仅在自学习策略能够对规则策略的性能进行提升时，激活自学习策略。因此混合策略兼具规则策略的可靠、可解释性和自学习决策在面对高维不确定性问题的优越性。该混合决策方法在具有强随机性的高速下匝道场景与高速安全驾驶场景中进行验证。测试结果表明，混合决策算法能够提升现有规则决策算法的表现。此外，为使混合决策避免由于训练数据不足导致的策略失效，本研究提出了有限训练数据情况下的混合决策可靠性保障方法，解决了在训练数据无法充分覆盖全部场景的情况下，未充分训练的混合决策策略失效的问题，使无人驾驶混合决策方法可以用于各种随机复杂交通场景。为使混合决策能够适应不同的地理特性，本研究提出了适应地理空间特征的决策场景构建方法，解决了无人驾驶汽车在变化的地理空间行驶过程中的决策自适应调整问题，使无人驾驶混合决策策略能够自动适应变化的地理空间。为使得该混合决策算法能够在实车平台上使用，本研究设计了集成混合决策的全栈自动驾驶算法平台，开展了大量的开放道路实车实验，实验结果能够验证混合决策的算法表现，实现了规则决策系统和自学习机制的有效融合，提高交通效率的同时保证安全。总的来说，本文提出的混合决策方法具有自学习能力，并且能够提升规则策略的表现，为现阶段无人驾驶汽车决策表现的进一步提升打下了重要的基础。

Motion planning is crucial in autonomous vehicle systems. It is a challenging problem when driving in high-dimension and uncertain traffic scenarios. Principle-based planning methods may fail due to the strong uncertainty. Learning-based methods can learn from past failures and use the neural network to adapt to the uncertainty, which have potential to improve the performance of the automated vehicles. However, the self-learning policy is unexplainable and may make some unexpected decisions. The intersection, between statistical machine learning and solid robotic system integration, is the bottleneck when using the self-learning technology in reliable autonomous vehicle systems. This work proposed hybrid planning methods, using both principle- and learning- based strategies, for automated vehicles. The hybrid planning can use the principled-based strategy to guarantee the lower bound, in the meantime, learn by itself during driving.In details, this work first proposed the hybrid planning methods, where the integrated principle- and learning-based strategy generate the trajectories synchronously. The learning-based strategy will be activated by an activation function. The activation function is designed to find whether the learning-based strategy can improve the principle-based strategy. Namely, the principle-based method guarantees the performance lower bound of the hybrid planning strategy, which also has self-learning ability. This method is implemented in highway-exit scenarios and collision avoidance scenarios in strong uncertain highway environments. The hybrid planning can improve the principle-based methods in both scenarios. Hybrid planning may fail when the training data cannot cover all the traffic cases. Therefore, this work further proposes the confidence aware hybrid planning method under the limited dataset. The confidence aware hybrid planning considers the distribution of the training data. The self-learning strategy only works when it has high confidence. Hybrid planning should be compatible with the road structures and local traffic characteristics. This work proposed the adaptive hybrid planning methods, which using the scenario templates to reconstruct the road structures. In this way, the driving policy can be transferred using the self-driving map. For real application, the work builds a full-stack driving platform, installing the hybrid planning method. The platform is tested in the traffic simulator and on the real vehicle. The results show that the autonomous vehicle can use the hybrid planning to drive and react the dynamic obstacles. In general, the proposed hybrid planning method has self-learning ability, and can improve the principle-based strategy. This technology can further improve the automated vehicle performance in the real road.