This work encomapsses the development of a robotic microdosing system. Microdosing is the science of precise delivery of powders into a chosen vessel. The system consists of two parts: loading material onto the dosing spatula and microdosing said material into the dosing dish. Both actions are exectued by following human demonstration.
The demonstration of the loading motion was recorded by kinesthetic guidance of the robot arm. The motion was parametrized with dynamic movement primitives (DMPs), allowing us to dynamically adjust the loading point in the loading area. The loading point's location is determined by scanning the powder surface with a laser scanner. The demonstrations of the dosing motion were captured using an optic motion tracking system. We adjusted the motion to be suitable for execution with the robot and parametrized it using periodic dynamic movement primitives (pDMPs). Microdosing is realised as regulation of the dosing motion's frequency, based on a model of the powder in conjuction with the measured mass in the dosing dish and on the spatula. We also considered and tested PID regulation, which proved unreliable.
We performed an evaluation of the system in the range [0,~30]~mg, using microcrystal cellulose (MCC) powder. The expanded uncertianty of the results is similar to that of measurements: 2.6~mg. The most significant contributors to the latter are the variablity of samples and environmental noise, predominantly changes in air pressure. Because the quality of measurements directly affects the quality of our model, we presume considerably better results could be obtained in a more suitable environment. The robot achieves greater precision than a human using the same technique. It's precision is comparable to that of trained experts, with some variance due to differing levels of experience. It's unable to meet the precision of purpose built machinery, falling short by about a factor of 10.
|
