search

moorepants / ncsrr-visiting-scholar-2014

Proposal materials for the Stanford NCSRR Summer Visiting Scholar Program in 2014.
0 stars 2 forks source link

Add speed ups #5

Closed moorepants closed 10 years ago

moorepants commented 10 years ago

Not allowing muscle activation to go below a certain value, for instance, and widening the active force-length curve. Softening the ground contact.

moorepants commented 10 years ago

You're altering the dynamics but you can still go back to the real dynamics when you are near the solution and not many more simulations are needed. I believe that CMA is robust enough that it will just keep going even if you alter the problem in midstream.

moorepants commented 10 years ago

Evaluation of muscle dynamics could be vectorized, because all muscles use the same equations, only the parameters are different, and they are independent.

moorepants commented 10 years ago

There is a lot of potential for using GPU in trajectory optimization with collocation methods. You evaluate the system dynamics at many nodes, this can happen at the same time. The major bottleneck however is in the solver. IPOPT can use a GPU-aware Pardiso solver.

moorepants commented 10 years ago

LBFGS is an unconstrained gradient-based method. I met Igor last year at Dynamic Walking. His work is unique in that it does not enforce the system dynamics as a hard constraint. The violations of system dynamics are computed via a collocation scheme and added as penalty terms to the cost function. It seems to converge very well even for complex movement tasks.

In engineering we frown upon violating Newton's laws, even a little bit, but for graphics it's probably fine.

I think Igor's approach could generate a good initial guess, and then you do another pass with a hard constrained optimization (SQP, SNOPT, or such) to get the final answer.