Deterministic planning for OMPL

[tlpss]

would be nice to have a convenient way to fix the seed (or maybe have this as default) so that generated plans are deterministic

[Victorlouisdg]

I tried this:

from ompl import util as ou

ou.RNG().setSeed(1)

But that gives the following error, even when you run it in the first cell of a notebook.

Error:   Random number generation already started. Changing seed now will not lead to deterministic sampling.
         at line 78 in [/project/src/ompl/util/src/RandomNumbers.cpp](https://file+.vscode-resource.vscode-cdn.net/project/src/ompl/util/src/RandomNumbers.cpp)

[Victorlouisdg]

Ok some progress, you can set the seed of the RNG object in its constructor:

However, I don't know yet if and how I can change the RNG object used by the SimpleSetup. I've found that StateSampler objects have a protected .rng_ attribute. And I've also found how you can allocate a new ValidStateSampler:

I'm still missing two things:

• How I configure the SimpleSetup to use the new ValidStateSampler
• How I change/set the RNG that ValidStateSampler uses

Additionally, there is a Python example on how to inherit from ValidStateSampler.

[Victorlouisdg]

A bit more progress, apparent

• RRTConnect does not allocate a ValidStateSampler, but simply a StateSampler, see RRTConnect.cpp.
• There exists a RealVectorDeterministicStateSampler

I've tried setting the allocator of the StateSpace use by my SimpleSetup like so:

def allocDeterministicStateSampler(space: ob.StateSpace) -> ob.StateSampler:
    return ob.RealVectorDeterministicStateSampler(space)

space.setStateSamplerAllocator(ob.StateSamplerAllocator(allocDeterministicStateSampler))

But I'm still not getting deterministic plans. When I create two different SimpleSetups and plan to the same goal, I still get two different paths. Maybe it's due to the path simplification also being random?

[zkytony]

Any updates on this issue?

[Victorlouisdg]

Hi Kaiyu, no we haven't had the time to figure this out yet, but this definitely still something we want in the future for repeatable testing.