Do all the three methods listed guarantee the convergence?
mylinhvu11
What is the classification of the function being great or not great? Is is the smoothness of the kinematic animation. Also, for te Gradient Decent, by being easy with ML tools, it seems more applicable and easier to implement, but the quality of the end result isn't displayed so does this method produce better results in representing inverse kinematics?
ld184
for gradient descent, convergence differs with different step sizes. How are the hyperparameters chosen here for this purpose?
alvin-xu-5745
Hyperparameters are flexible and have many ways to be chosen, just like with any other ML model that is not necessarily graphics-related. Things like step size can be determined experimentally or in slightly more analytical methods.
Here is a cool demonstration video showing gradient descent used in practice for inverse kinematics. It is clearly visible that it is slower than analytical methods, though it is much more flexible.
https://youtu.be/udUTlp45Oqo
Noppapon
One of the applications for Inverse Kinematics is in robotics, where multiple robot parts connected by joints are involved. Hence, in some cases, we might not only want to know the orientation of each joint for the robot's hand to reach the desired destination, but also how to do so most effectively (required least time/least movements).
Do all the three methods listed guarantee the convergence?
What is the classification of the function being great or not great? Is is the smoothness of the kinematic animation. Also, for te Gradient Decent, by being easy with ML tools, it seems more applicable and easier to implement, but the quality of the end result isn't displayed so does this method produce better results in representing inverse kinematics?
for gradient descent, convergence differs with different step sizes. How are the hyperparameters chosen here for this purpose?
Hyperparameters are flexible and have many ways to be chosen, just like with any other ML model that is not necessarily graphics-related. Things like step size can be determined experimentally or in slightly more analytical methods.
Here is a cool demonstration video showing gradient descent used in practice for inverse kinematics. It is clearly visible that it is slower than analytical methods, though it is much more flexible.
https://youtu.be/udUTlp45Oqo
One of the applications for Inverse Kinematics is in robotics, where multiple robot parts connected by joints are involved. Hence, in some cases, we might not only want to know the orientation of each joint for the robot's hand to reach the desired destination, but also how to do so most effectively (required least time/least movements).