API question: optimal control
Posted: Sun Mar 13, 2011 2:29 pm
Hi all
I encountered troubles while trying to use API to conduct an optimal control problem.
API codes were implemented in order to find a set of muscle activation which could produce maximal linear velocity of foot segment.
The initial and final muscle activations of leg acceleration phase are used as initial guess, and the objective function is the value of foot velocity along X axis.
The results are always veered away from real human like motion. The solution attempt to accelerate knee joint immediately, rather than follow the proximal-to-distal pattern, which means accelerate hip joint first, followed by the acceleration of more distal segments.
Please take a look at these two video clips:
http://www.youtube.com/watch?v=vmo7FFakfDY
http://www.youtube.com/watch?v=Qq6kCm5w_Jc
I tried to tweak objective functions, adding hip joint angles or joint angular velocity as penalty function. But the results still couldn't reproduce performance as fast as human subjects.
Is it possible to lead IPOPT algorithm to the correct direction to search for optimal solution?
Any comment would be appreciated.
Thanks!
I encountered troubles while trying to use API to conduct an optimal control problem.
API codes were implemented in order to find a set of muscle activation which could produce maximal linear velocity of foot segment.
The initial and final muscle activations of leg acceleration phase are used as initial guess, and the objective function is the value of foot velocity along X axis.
The results are always veered away from real human like motion. The solution attempt to accelerate knee joint immediately, rather than follow the proximal-to-distal pattern, which means accelerate hip joint first, followed by the acceleration of more distal segments.
Please take a look at these two video clips:
http://www.youtube.com/watch?v=vmo7FFakfDY
http://www.youtube.com/watch?v=Qq6kCm5w_Jc
I tried to tweak objective functions, adding hip joint angles or joint angular velocity as penalty function. But the results still couldn't reproduce performance as fast as human subjects.
Is it possible to lead IPOPT algorithm to the correct direction to search for optimal solution?
Any comment would be appreciated.
Thanks!