Compared with machine learning model, physical model has some advantages:
It has a good generalization because we do not set any limit on predicted point. We can easily implement it again into another system.
The robot arm has a larger reachable set and we do not to store too many parameters.
However, the behavior of physical model is relatively worse than machine learning model, reasons are as follows:
The prediction point is not very accurate. We neglect air resistance and assume the gravitational acceleration as 9.81, which is not always the real situation. Second, the image we captured has some distortion and the camera need to be calibrated. Also, we measure the transformation between coordinates by hand and there exists some error in mapping.
The move maybe unsafe because the result is hard to control. Sometimes it gives unreachable position to controller. Also, sometimes the built-in function to calculate the inverse kinematic is unstable.
This model has a high requirement on device quality. To get the best result, a high-frequency-camera is needed.
Machine learning model
Analysis
Compared with physical model, machine learning model shows a better performance. Here are the advantages of this model:
If given enough data, If given enough data, it can predict the dropping point faster and more accurate. In our project we collect 284 sets of data.
Because all predictions lie in the convex hull of training data. So the robot arm would not move to a unreasonable position, and it is safer and more stable.
Because we do not need to calibrate camera or measure any mappings. The complexity of system is reduced.
However, there still exists some shortcomings:
It costs a lot of time to collect data.
According to the model, the robot arm would not move out of the configuration we set. So it cannot deal with the trajectory out of reachable set.
It is hard to implement again on other ball-catching system as the type of robot arm and position of camera may change.