Visiegebaseerde autonome en op afstand bestuurde robotinterventies voor de farmaceutische industrie

Modern medicine has allowed humanity to combat and cure life-threatening diseases. Some of the most dangerous health problems of the past have been fully eradicated thanks to the pharmaceuticals of today. This in turn has improved both life expectancy and the quality of life around the globe. Because of this, the most important aspect during pharmaceutical aseptic production is the assurance of aseptic excellence and product quality, forcing companies to stay true to often outdated but validated techniques. High-speed batch-based medicine production is still performed in glove port-equipped isolator cells, where human operators are responsible for the handling of interventions within the process. However, the outcry for innovation from within pharmaceutical companies is becoming louder and louder. The required time, cost, and hazard of contamination related to the inspection, maintenance, and validation of the gloves and their manual interventions drastically impedes with the runtime efficiency of the production lines. Ultimately, it is generally accepted that the future of pharmaceutical production should become completely gloveless. One promising way to achieve this is by the use of robots within isolators, where they become responsible for the execution of interventions in either an autonomous or teleoperative manner. This work contributes to both the future of aseptic isolator-based production, and robotics in general, as the majority of the proposed work could be applied in various fields. The first contribution corresponds to an optimisation method used to find more consistent and more optimal robot trajectories in point-to-point sequences when using collision-free trajectory generation algorithms such as Probabilistic Roadmap (PRM) and Rapid Random Tree (RRT) planners. These algorithms are known to often return overly long and complicated trajectories due to their probabilistic nature. Optimisation is achieved by simultaneously planning and filtering trajectories based on the shortest path duration while the robot is executing a previous motion of the sequence. Simulated experiments have shown an increased consistency of the found trajectories over multiple planning attempts of the same pose sequences. A second contribution was made in the field of task management. Here, each known intervention that could occur is seen as an individual task structure. In a realistic scenario, many of these tasks could be required at the same time during production. The proposed Behavior Tree-based management method addresses this by taking an object-oriented approach. When a task should take place, the manager creates a unique task event object for it tied to a priority level, housing all relevant information regarding the task within. With this, the manager can autonomously sort, prioritise, pause, cancel and (re)start interventions as required. Additional logging functionality was added for human readability and traceability. The third and fourth contributions are both centered within teleoperation. For the third, a general sensorless (haptic) safety system was implemented that uses the input of the user to calculate and anticipate the future robot state, and check it against the state of the robot and the known environment. Both singularities and collisions are prevented by applying a push-back velocity to the robot based on the input command. For singularity prevention, the determinant of the Jacobian matrix is used, while for collision prevention, the normal direction of the closest collision geometry is extracted via ray casting. Safety zones can be assigned by the user in which select parts of the robot can enter and move under additional constraints. In the fourth contribution, this work was further specialised to serve the plunger removal intervention. Here, a YOLOv8-based plunger detection algorithm signals the intervention, after which the operator can remove the anomaly teleoperatively. The movement of the robot is strategically locked to make the control as intuitive for the operator as possible, combined with haptic feedback to simulate the feeling of the lane of the plunger. This method could be performed in a static and dynamic setup, either locally (the operator/teleoperation interface have a direct connection to the main computer) or fully remote (the operator/teleoperation interface are separated from the main computer via a remote network configuration). Both contributions were tested with unique user groups, which have indicated a high level of intuitiveness for both methods. To conclude, a practical contribution to the open source robotics community can be found in the development and maintenance of a ROS2 driver for the used Stäubli robot. This driver allows the use of state-of-the-art collisionfree trajectory planning algorithms, real-time pose and velocity-based jogging control, and direct IO access. It also adds the possibility to add up to four synchronously driven external axes adding up to four degrees of freedom to the system.