In these 3 benchmarks, the constraint graph is built in python. When two neighbor edges are foliated, a waypoint edge with a level set edge at the first place is inserted in both ways.
Let us rank these benchmarks in increasing order of constraints:
- case 1: 2020-04-10,
- case 2: 2020-04-11,
- case 3: 2020-04-09.
The differences between these cases resides in the following constraints
- grasp/complement between waypoints pregrasp and intersect,
- placement/complement between waypoints preplace and place. In this case, locked joints are not used as the concatenation of placement and placement/complement (factory.constraints.strict = True)
| Case | grasp/comp | placement/comp |
|---|---|---|
| 1 | no | no |
| 2 | yes | no |
| 3 | yes | yes |
| Benchmark | case 1 | case 2 | case 3 | |
|---|---|---|---|---|
| baxter-manipulate-boxes | 4 | 3.6 | 15.8 | time (s) |
| 153.4 | 133.99 | 306.1 | nodes | |
| construction-set | 10.9 | 3.71 | 10.71 | time (s) |
| 47 | 39.4 | 23.85 | nodes | |
| pr2-manipulation-kitchen | 17.25 | 15.78 | 62.45(*) | time (s) |
| 128.5 | 121.6 | 452.5 | nodes | |
| pr2-manipulation-two-hand | 4.24 | 4.43 | 22.19 | time (s) |
| 44.75 | 40.6 | 315.45 | nodes | |
| romeo-placard | fail | 155.85 | 180.56 | time (s) |
| fail | 712.0 | 819.35 | nodes | |
| ur3-spheres | fail | 67.68 (**) | 8.49 | time (s) |
| fail | 3482.0 | 260.3 | nodes |
(*) Note that in case 3, benchmark pr2-manipulation-kitchen where there are two contact surfaces, along solution paths, the object discontinuously jumps from one surface to the other one. (**) It turns out that level set edges going from a grasp state to state "free" always fail due to the fact that locked joints are used as complement of placement constraints. This leads to inserting incompatible preplacement and locked joint constraints in the target constraints.
- There might be a problem with level set edges inserted in waypoint edges. Enforcing the same right hand side along each edge of the waypoint edge for the parameterizing constraints may result in a difference bigger than the error threshold and therefore to a configuration that is considered as on a different leaf as expected. This may explain the poor result of ur3-spheres.