[RSDK-13624] Introduce support for randomized trajectory testing#192
Conversation
| BOOST_CHECK_EQUAL(segments.size(), 4); | ||
| } | ||
|
|
||
| // Deduplication must compare against the last kept point, not the |
There was a problem hiding this comment.
We had a bug in the trajex waypoint deduplication code! I found it sort of by accident while looking at some of the random trajectories, but I'd like to fix it here and now.
| if (xt::norm_linf(waypoints[i] - waypoints[i - 1])() > tolerance) { | ||
| result.add_waypoint(waypoints[i]); | ||
| std::ranges::for_each(waypoints | std::views::drop(1), [&](const auto& waypoint) { | ||
| if (xt::norm_linf(waypoint - result.back())() > tolerance) { |
There was a problem hiding this comment.
If you look carefully, you will see that the old loop only compared against the prior waypoint, not the most recently kept waypoint.
| // Ensure the original last waypoint is preserved exactly as the destination. | ||
| // If it was deduplicated away, replace the last kept point with it. If the | ||
| // entire sequence collapsed to a single point, leave it as the start and let | ||
| // path creation fail with a clear error. |
There was a problem hiding this comment.
This seems like an obvious and desirable improvement, in that we really do want the path to start and end at the first and last waypoints.
| return result; | ||
| } | ||
|
|
||
| // Expected number of reversals and colinears per random run. The per-step injection |
There was a problem hiding this comment.
There is no need to review this test infrastructure carefully unless you are interested.
|
|
||
| // Ensure that if we run the random trajectory twice on the same seed, | ||
| // that we get deterministic outcomes. | ||
| BOOST_AUTO_TEST_CASE(random_trajectory_determinism) { |
There was a problem hiding this comment.
This is the only test that gets lit up as part of this PR: it generates two trajectories from the same random seed and checks that we get the same events. Trajectory generation should be fully deterministic, so this is a good check to have.
2 participants
This adds support for, but does not enable, randomized testing of trajectory generation. It's found some interesting things already, but we have higher priorities. However, since we are shortly getting trajex into it's own repo, I'm trying to get anything I have on a branch out for PR so I can get it merged ahead of the split.
I'll leave some comments on what is going on in various places.