search

hanyas / trajopt

A toolbox for trajectory optimization of dynamical systems
MIT License
50 stars 16 forks source link

do you have all versions of gps's implementation #2

Closed dujinyu closed 4 years ago

dujinyu commented 4 years ago

do you have all versions of gps's implementation?

hanyas commented 4 years ago

What do you mean? What versions are you referring to?

dujinyu commented 4 years ago

Guided Policy Search(GPS) Variational Policy Search via Trajectory Optimization(vGPS) Learning Complex Neural Network Policies with Trajectory Optimization(cGPS) I mean the implementation of the above three papers

hanyas commented 4 years ago

No this repository only implements the core of GPS, what is referred to as called Maximum Entropy iLQG. The iterative trajectory optimization via dynamic programming.

dujinyu commented 4 years ago

Is this repo based on an article you published? I find some new abbreviations, such as elqr, bspilqr.

hanyas commented 4 years ago

No, the implementations are not really related to a publication directly. Although you can find more explanation on Max-Entropy iLQG here: https://www.ias.informatik.tu-darmstadt.de/uploads/Site/EditPublication/Abdulsamad_ICAPS_2017.pdf

And here: https://www.ias.informatik.tu-darmstadt.de/uploads/Site/EditPublication/Abdulsamad_Master_Thesis_2016.pdf

The names eLQR and bsp-iLQR stand for Extended LQR and Belief-Space iLQR by Jur Van den Berg: https://www.researchgate.net/publication/258140864_Motion_Planning_under_Uncertainty_Using_Iterative_Local_Optimization_in_Belief_Space

Also, if you are interested in robust trajectory optimization techniques check our latest paper on the subject: https://www.ias.informatik.tu-darmstadt.de/uploads/Team/JoeWatson/Watson19I2c.pdf

dujinyu commented 4 years ago

Thank you very much for your reply. I know a little about Extend LQR. elqr method needs the dynamics is known. if the dynamics is not known, how can i use elqr, when the dynamics is modeled by sampling.

hanyas commented 4 years ago

In the case of unknown dynamics you can use these methods either by fitting a global model and linearizing it around a given trajectory or by directly fitting time-variant linear dynamics at each time step.

Now if you don't mind I will be closing this issue, because this is not purpose of this interface