Autonomous Driving Control With Closed-Loop Decision and Planning

The patent merges the behavior decision-making layer and motion planning layer into a unified optimization framework. Both layers are optimized simultaneously using a joint loss function that incorporates objectives from both layers, allowing them to work together as an integrated system rather than separate modules. This combining approach enables improved safety performance while managing system complexity through unified training.

The patent introduces a reward model as an intermediary component that evaluates and guides the optimization of both the behavior decision-making layer and motion planning layer. This reward model acts as a mediator that provides feedback signals to both layers, enabling coordinated optimization without directly increasing the complexity of either individual layer. The reward model bridges the two layers and facilitates their joint improvement in safety performance.

The patent implements a feedback mechanism where a reward model evaluates the outputs of both the behavior decision-making layer and motion planning layer. The reward model provides feedback signals that guide the optimization of both layers, allowing the system to learn from evaluation results and continuously improve safety performance. This feedback loop enables the hierarchical architecture to achieve optimal performance by iteratively refining decisions based on evaluated outcomes.

The patent makes the optimization process dynamic by using a reward model that can adaptively evaluate different scenarios and provide context-dependent feedback. The reward model allows the system to dynamically adjust its optimization priorities based on the specific driving situation, enabling the hierarchical architecture to achieve better safety performance across diverse scenarios rather than relying on static optimization parameters.