The invention discloses an end-to-end automatic driving decision planning method and device combined with meta-learning multi-task optimization, an end-to-end learning mode is adopted to overcome error transmission, a multi-task constraint method is utilized to optimize model convergence, and multi-branch output results of multiple tasks further increase the interpretability of the model. Eight different sub-task branches are designed, including environment perception tasks and self state cognition, and original sensor features are converted into aerial view space through a sub-task design model, so that the accuracy of decision planning prediction is improved. Furthermore, in order to optimize different multi-task combinations, firstly, the multi-task intimacy is adopted to evaluate the different task combinations to obtain initialized task weights, and then a multi-task optimization method based on combination of meta-learning is adopted to dynamically adjust the weights among the different tasks in the training process, so that optimization of main task prediction is realized.
本发明公开了一种结合元学习多任务优化的端到端自动驾驶决策规划方法及设备,采用端到端学习方式克服误差传递,同时利用多任务约束的方法优化了模型的收敛,多任务的多分支输出结果进一步增加了模型的可解释性。设计八个不同的子任务分支,包含对于环境的感知任务以及对于自身状态的认知,通过子任务的设计模型将原始传感器特征转化到鸟瞰图空间提升了决策规划预测的准确性。进一步,为优化不同的多任务组合,首先采用多任务亲密度对不同任务组合进行评估,得到初始化任务权重,然后采用基于结合元学习的多任务优化方法,在训练过程中动态调整不同任务之间权重,从而实现主任务预测的最优化。
End-to-end automatic driving decision planning method and device in combination with meta-learning multi-task optimization
一种结合元学习多任务优化的端到端自动驾驶决策规划方法及设备
2023-09-12
Patent
Electronic Resource
Chinese
European Patent Office | 2024
|European Patent Office | 2023
|Automatic driving planning decision-making method and device
European Patent Office | 2024
|European Patent Office | 2024
|Multi-task sensing method and device for automatic driving
European Patent Office | 2024
|