机器人散乱件分拣中的深度图像特征迁移

Feature Transfer of Depth Image in Bin-picking

在当前阶段机器人操作的各类研究方法中,都离不开仿真环境的辅助,而将仿真环境中得到的方法应用于真实环境中也成为一项具有挑战性的任务。针对机器人无序分拣场景,使用CycleGAN方法,进行仿真环境与真实环境中深度图像的特征映射学习,在真实环境深度图像上附加仿真环境中的深度图像特征,提高在仿真环境中训练得到的方法在真实环境中的鲁棒性。分别使用自建数据集以及WISDOM数据集对这一方法进行了验证,研究了特征映射过程中各变量对映射器学习结果的影响。

At current phase,research methods for robot operation are inseparable from the assistance of simulation environment,and the application of methods from simulation environment to real environment has become one of the challenging tasks.Focused on robot bin-picking,in this paper,CycleGAN is used to study the feature mapping of the depth image between simulation environment and real environment,and then attach depth features in simulation environment to depth image in the real environment.Such an operation could improve the robustness of the method trained in the simulation environment when deployed in the real environment.And the influence of variables in the mapping is studied.