Choreo Path Generator 26.0.3 for Windows generating VERY bad path

[msonnier]

Path TrenchLeft2 generated properly on prior version, but with 26.0.3 it does generate a path after a VERY long time, but the generated path is totally broken. Diagnostics file attached.

choreo-diagnostics-DKTN4CUl.zip

[calcmogul]

The trajectory successfully generates on 89574f9 but has a line search failure on the commit after that: 57d281b (angular acceleration kinematics fix).

On main, the solver prematurely reports success because some NaNs showed up. I need to trace the source of those.

EDIT: I'm getting inconsistent results for each solve. Could be memory corruption or an uninitialized read somewhere.

[witherrrr]

It solves just fine with a slightly lower dt step - from some manual testing, it seems that the very short distance between Waypoint 4 and Waypoint 5 isn't handled well by the solver (for very short segments the control interval heuristic will not produce sufficiently many segments for a good solve). We can fix this class of issue directly by either

• Adding a dt cap based on minimum distance between waypoints, similar to the anti-tunneling logic we already use to cap dt to handle keep-out constraints. This would raise the size of the total problem a decent bit but may be more reliable.
• Putting a lower bound on the number of control intervals on each segment. Should directly fix instances like these without too much added work.