Retuning, reduces avg time by almost a second

[glpuga]

Tried over 50 episodes. There was a single collision, but since I can't know the start seed I don't know the reason for the collision.

Looks promissory, though, would reduce the lap time to between 9.7 and 9.9 seconds.

[glpuga]

[glpuga]

Put this on hold until further notice, found a problem...

[glpuga]

Looking good again.

I needed to help it with your "minimal corrections" commit to get it through the finish line.

In 100 episodes there were two collisions on the second gate. I've got no record of the start seeds, but in that gate I suspect that the minimum correction may have hit the frame when turning left.

If we have time, check if you can verify the performance of this change.

[hidmic]

@glpuga Indeed, we get a nice reduction in average flight time:

But we have lots of collisions:

• Hitting the gate frames (seeds 1303, 1320, 1323, 1337, 1346, 1352, 1361, 1371, 1375, 1395, 1396). I think the problem here is that we're not rolling/pitching enough. At these speeds and this close to the gate frames, we can only avoid the collision by approaching the gate at an angle. Constraining maximum inclination to 30 degrees, reducing body rates weights to 1 (or less), and extending the correction sigma to 0.75 m seems to fix the issue.

  https://user-images.githubusercontent.com/13500507/196043938-b1f9c03f-7aae-464a-af47-37d6cdce745c.mp4

• Hitting the floor when destabilizing at the start (seeds 1380, 1382). Not sure about these ones.

[glpuga]

:cry: well, it was a good try, closing this.

[hidmic]

@glpuga I think this is promising, even if we decide not to merge for the competition. It just needs some tuning.

[glpuga]

I agree, but tuning is time consuming, so I think at this stage it's better begin closing the submission and just test that what we submit is not missing anything important.

10.8 sec is good, especially if it requires no training, and it's more robust. We'll do fine.