复杂海洋环境下无人艇自主感知与决策关键技术

In many cases, the undefined and incomplete information collected from complex Marine environment can affect the self-perception and decision making system of ( unmanned Surface robot) USV and cause the low mission completion rate and poor dynamic environment adaptation ability. In order to solve this issue, the key technology of autonomous perception and decision making of USV is proposed, which is based on the principle of cognitive process of brain memory. By simulating the functional modules of memory units and the relationship among these modules in the cognitive process of brain memory, the framework of the autonomous perception and decision-making system of USV is proposed. By the off-line learning of large samples and on-line identification and learning of small samples, the framework of functional modules of memory units can be simulated and realized, as a result, USV owns the priori knowledge of target recognition, judgement of intention and situational awareness in complex Marine environment. Based on the activation diffusion, selective attention and cross-perception of information among functional modules of memory units, we explore the model and method of self-perception and decision making system of the USV; and then, a set of models and methods can be formed, which is about how the USV can complete the self-perception and decision making in complex Marine environment, therefore, the problem of low mission completion rate and poor dynamic environment adaptation ability of USV in complex Marine environment can be resolved; Finally, by building simulation platform for realistic test, we provide the technical specification, which is not only for promoting and using this method, but also for providing essential guarantee of that USV will play a key role in complex sea areas and in the military field.

针对复杂海洋环境不确定与不完整信息给海面机器人即无人艇自主感知与决策带来的任务成功率低和动态环境适应能力差等问题，提出了基于脑记忆认知过程的无人艇自主感知与决策关键技术。通过模拟脑记忆认知过程中各记忆单元功能模块及其关系，提出无人艇自主感知与决策框架；通过大样本离线学习及小样本在线识别与学习，以模拟和实现该框架中各记忆单元功能模块，形成复杂海洋场景中目标识别、意图判定及态势感知的先验知识；基于记忆单元各模块间信息的激活扩散、选择性注意和交叉感知等机理，研究无人艇自主感知与决策的模型与方法；继而，可形成一套在复杂海洋环境下无人艇自主感知与决策的模型与方法体系，以解决无人艇在复杂海洋环境下任务成功率低和动态环境适应能力差等问题；最后，通过构建仿真平台与实艇验证，为本方法推广与使用提供技术规范，及为无人艇在复杂海域和军事领域充分发挥杀手锏作用提供重要保障。

针对复杂海洋环境不确定与不完整信息给无人艇带来的任务成功率低和动态环境适应能力差等问题，本项目形成一套在复杂海洋环境下无人艇自主感知与决策的模型与方法体系，基于搭建的仿真与试验平台对方法进行了验证。项目的主要研究结果包括：.（1）基于脑记忆认知过程的无人艇自主感知和决策框架。结合短期情景记忆中目标信息和长时记忆知识表达实现知识补全；通过跨模态信息关联与融合实现目标检测与意图判断，通过域适应方法提升感知模型泛化能力，并与环境互动生成反馈实现自主决策，以应对多变海洋场景中信息不确定与不完整对自主感知与决策带来的影响。.（2）复杂海洋环境下目标实体及场景知识的表达和补全。提出基于“蒸馏学习”和“联合学习”的知识表达与补全方法，通过将二维图像先验信息引入激光点云数据实现动态目标及场景知识表达；提出基于时空上下文场景补全语义分割方法，引入长时记忆先验信息补全稀疏残缺激光点云。.（3）基于动态贝叶斯网络和深度Q网络无人艇自主决策方法。通过对无人艇威胁评估以及与环境之间交互与反馈，在环境中探索并收集信息，将收集的信息保存于“长时记忆”之中，形成一种脑记忆认知过程，从而实现无人艇在动态海洋环境中的自主决策。.（4）复杂海洋场景无人艇自主感知与决策仿真与实艇验证平台。利用物理引擎对无人艇动力学、传感器、海洋环境等信息进行模拟，模拟出信息不正整和不确定下的复杂海洋场景，基于仿真平台进行大样本离线学习。.项目探索的关键技术可以有效降低信息不确定与不完整对无人艇任务执行率和环境适应性的影响，实现复杂海洋环境下无人艇实时与准确的自主感知与决策，为海洋无人系统的自主化与智能化提供了理论基础与技术支撑。