comparing the planners in openrave and moveit

[Karthikeyan Yuvraj]

Hello Everyone,

I have been looking into planners used by openrave and moveit. I was told that moveit uses a SBPL planner called KPIECE and Openrave uses RRT based planner.  SBPL planner KPIECE is a kinodynamic planner and it is widely used in willow garage(am I right?) and why is it so?. Is it better than using RRT or RRT*?. Also among RRT based and SBPL based KPIECE which is faster?. Thanks

regards,

Karthik.

[Ioan Sucan]

Hello,

Moveit uses a plugin system for connecting different plannig libraries. One of these planning libraries is OMPL, a sampling based motion planning library. OMPL includes KPIECE + a few variants of KPIECE, as well as RRT and a few variants of RRT, including RRT*. We can also use the SBPL library, but the bindings for that need a bit of work. We do not use kinodynamic planners at willow garage. We use default settings from OMPL when the user does not specify the planner to use explicitly. In that case, a kinematic, lazy bi-directional version of KPIECE is used. That is usually faster than RRT. Please see ompl.kavrakilab.org/planners.html for a complete list of planners in OMPL.

Ioan

[Karthikeyan Yuvraj]

Hello Ioan,

Lazy bi-directional version of KPIECE is used as default in OMPL(so I assume this is used in moveit too when I don't specify the planner),but I read your paper on KPIECE http://ioan.sucan.ro/files/pubs/wafr2008.pdf it starts by saying that this is a kino-dynamic motion planner. I am confused ,Is lazy bi-directional KPIECE is different than KPIECE?. 

Also Is there a reason why Willow garage is inclined to use  lazy bi-directional version of KPIECE or a varient of KPIECE?

karthik

[Ioan Sucan]

Hello Karthik,

You are correct, the initial version of the planner was introduced as a planner for systems with kinodynamic constraints (implemented in OMPL as control::KPIECE). The same exploration strategy can be applied when planning with simpler geometric constraints only. (This is implemented in geometric::KPIECE). Typically, sampling-based planners perform better when there are two trees being grown: one from start and one from goal. This is the bi-directional part (implemented as geometric::BKPIECE). Additionally, lazy collision checking can be used (geometric::LBKPIECE). (For some more details on these techniques, please see http://ioan.sucan.ro/files/pubs/implementingplanners_icra2010.pdf) The choice in planner is frankly an arbitrary one, based on what seemed to work most of the time; It is likely different planners will perform better for some problems. For instance SBL and RRTConnect perform quite well.

The choice is really an if statement:

 - if a projection was specified by the user in ompl_planning.yaml, then use LBKPIECE;

 - otherwise, use RRTConnect.

This is certainly not a choice that will work best in all cases, but has performed reasonably so far. Please let us know if you are you encountering problems, and perhaps we can improve this planner selection strategy.

Ioan

[Zhang Jun]

Hello Loan and Karthik,

I am wondering how to use a kinodynamic motion planner under Moveit?  There are some kinodynamic planners under OMPL::control, but how to use it? There are some demo in OMPL about how to invoke a planner with control directly. But it seems there no example about how to invoke those planners through Moveit.

Thanks in advance
Jun

[Ioan Sucan]

Hello Zhang Jun,

Unfortunately MoveIt does not expose any of the control/ planners in OMPL. So planning in MoveIt with OMPL is currently limited to planning under geometric constraints.

Ioan