Snap mprims perfomance review

1. Num of Heuristics: 20
   Snap mprims: Full Body(radius 3x) and Base
   Tuck and untuck mprims disabled
   Result: Much worse. No improvement.

2. Heuristics: Default set
   Snap mprims: Full body (radius 3x for base_x_y and 15x for yaw)
   Planner: DTS Plus
   eps: 20
   Result: Worse than baseline dts_focal but restricting yaw improves wrt dts_plus.

3. Heuristics: Default set
   Snap mprims: Full body (radius 3x for base_x_y and 15x for yaw) and Base snap (radius 30) Planner: DTS Plus
   eps: 20
   Result: Worse than [2]

4. I tried choosing the start of a peak as center of activation circle and end of the peak as island state. However, curiously, the two states seem to be very close.

5. Snap mprims work better with plus and unconstrained planners. Increasing the yaw activation radius increases the number of failed collision checkings.

6. Things to try:

   • Keep number of islands low (even at the cost of higher activation radius).
   • Only use searches from training data which are close to my current search problem for calculating the island states.
   • Use snap primitives only for anchor search.

7. Base snap motion leads to large improvements. Perhaps, limit the use of full body snap? With 10 heuristics, specialized selection of snap mprims (for 0, 1, 2 heuristics) work best. Limit to anchor with 20 heuristics?

8. Current conclusions:

   • The combination of body, base and arm snaps works best.
   • Use of distance thresholds for mprims leads to issues when the goal is near a bottleneck. For example, baseSnap is deactivated when search is within threshold, similarly for base adaptive mprim.