A semi-autonomous approach for robotic grasping inside a glovebox is presented, as control is shared between the robot's planning and control algorithms, and the user's validation and selection inputs. This approach aims to improve task performance and flexibility by relying on the user's flexible planning and the robot's precise task execution. Poor visibility, clutter and uncertainty of objects inside glovebox environments make handling items very challenging. We focus on cases where solid debris need to be cleared from an area. This can be dangerous and cumbersome when attempting manually, as they could include sharp and hazardous items. While autonomous grasping models supported by machine learning techniques can be applicable in these scenarios due to its efficiency and accuracy, certain safety critical operations may require human input. In the semi-autonomous grasping system, we take advantage of the intermediate outputs of the generative grasp models. By using the generated grasp probability as a 2D map, multiple grasp probabilities are presented to the operator through a user-friendly interface, which allows them to direct the robot towards the appropriate direction.

Second fund from EP/W001128/1, Robotics and Artificial Intelligence for Nuclear Plus (RAIN+)