You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
precompute the dubin/reeds-shepp distances for use in heuristic shouldn't do since no static graph anymore on initialization
if exact increments of theta, we can precompute the getProjections for all bins. DONE
if exact increments of theta, we can precompute the footprint collision detection rotations. too involved in costmap 2d, would actually be useful, but out of scope for just this planner. Maybe something to consider for another time.
Parallelize
parallelize collision checking in analytic expansion -- too many atomic operations to test quickly and too rarely run
parallelize collision checking lines -- didn't help
parallelize 3 neighbor find/validate/expand (I dont think 3 is enough to parallelize and be effective with overhead)
General
find ways to minimize the number of nodes expanded (motion models move more? lower costmap resolution? hueristics drive better, etc) Heuristics very good at it and with the costmap downsampler, this is essentially as good as it gets!
profile code and see where we can make things faster.
Look at it with fresh eyes and see if we can simplify / make it better through more standard iterative approaches we have better control over. This currently seems a little rigid and doesn't do all that I'd like it to do (though might just be tuning)
These have been evaluated and precomputing the heading bins is in progress, closing because rest did not either make sense, impact performance enough to be worthwhile, or too time consuming for the marginal benefit. Path smoother ticket branched out into a new ticket.
Precompute
precompute the dubin/reeds-shepp distances for use in heuristicshouldn't do since no static graph anymore on initializationif exact increments of theta, we can precompute the footprint collision detection rotations.too involved in costmap 2d, would actually be useful, but out of scope for just this planner. Maybe something to consider for another time.Parallelize
parallelize collision checking in analytic expansion-- too many atomic operations to test quickly and too rarely runparallelize 3 neighbor find/validate/expand (I dont think 3 is enough to parallelize and be effective with overhead)General
find ways to minimize the number of nodes expanded (motion models move more? lower costmap resolution? hueristics drive better, etc)Heuristics very good at it and with the costmap downsampler, this is essentially as good as it gets!remove compute of heuristic twice in each node on visitation (e.g. Add waitForServer to TaskClient #45)these are on different nodesapproximating cos / sin / tan functionsSmoother:
The text was updated successfully, but these errors were encountered: