In this work, a novel local motion planning algorithm is presented, for underwater vehicle manipulator systems (UVMS) that perform autonomous underwater inspection operations. An optimization problem is formulated considering the collision avoidance, the approximation of the given task curve and critical optimization criteria. The searching method is based on an evolutionary algorithm and it is able to generate a local motion plan using continuously updated sensor information. The working environment is represented by a Bump-surface entity, constantly updated by a parallel algorithm implemented on a graphical processing unit (GPU). A trained artificial neural network is used for the fast approximation of the considered dexterity index. The local planner can cope with unknown obstacles inside the workspace while executing the task and pursuing high performance configurations in the free space. A welding inspection on an underwater tube structure is considered as the validation scenario, while a UVMS with a mounted six degrees of freedom manipulator is assigned to perform the task.