Flexible robotic arms have unique advantages in narrow or obstacle-rich workspaces due to their high flexibility and compliance.When a flexible robotic arm operates in a narrow

unstructured and complex environment

the use of vision for arm shape perception often suffers from occlusion and limited field of view

which results in the absence of arm shape backbone features in the input image

thus reducing the arm shape perception precision of flexible robotic arm.Aiming at this problem

a convolutional attention mechanism-based monocular shape sensing neural network model (AT-MoSSNet) is proposed.Ablation experiments are conducted on a public dataset.Compared with the baseline model

the proposed model reduces the mean error of robotic-arm shape(MERS)by 16.5% and the mean error of robotic-arm tip(MERT)by 29.8%.Finally

a random mask is added to the test set to simulate external occlusion of the flexible arm.The test results show that the MERS and MERT are reduced by about 16% compared with MoSSNet

which verifies that the proposed method can effectively handle situations such as field of view occlusion and complex background.