Looking for people with experience with the Coral + LL3, mostly asking whether the FPS usable and data consistent and usable. If there’s anything else which is notable about your experience with the Coral that would be great as well.
1 Like
iirc the provided note pipeline runs at about 17 fps, which is good enough for auto lineup if you do some trig to find the field position of the note and close loop on odometry. Not really good enough for close looping on the raw camera readings, especially for a turning based lineup. Also, depending on the camera mount angle you might have issues with not getting enough frames to align with while driving very fast. We ended up driving at half speed while auto intaking for consistency.
1 Like
Note detection was wonderful- very precise and the framerate we got was decent (15-30 fps, if I remember right). However, we ran into a problem where the entire device would frequently restart, disabling the ML pipeline on reboot. So, it was unusable outside of a testing environment. We think it was static discharge, but we’re unsure- most reasonable isolation methods weren’t very effective for us, and the problem occurred with numerous combinations of Limelights, Corals, and sufficiently specced cables. I have seen other people mention the same issue online, and as far as I know no reliable solution has been found 
3 Likes
I’d be curious to know the drop-off of the readings, if there was one for the pipeline.
We ran one for automatic note alignment and it actually worked too good a couple times. We had it set to take steering away from the driver to make sure he picked them up straight(he had a bad habbit of straffing accross the note which would make it jam in the intake) away we got stuck trying to pick up a note that had passed through the amp and would lie there on the other side of the lexan wall if the human player didn’t move it.
1 Like
what do you mean by drop-off?
For 2928, we ran note detection for intaking in both autonomous and teleop throughout the entire season. It wasn’t anything complicated, but whenever the intake button was held, it would strafe the robot to stay lined up with the note. This helped a lot when the note was not visible by the driver. It’s been a while, but i think the framerate was in the ballpark of 30 fps which was more than usable for us.
We had a few instances where the coral would momentarily disconnect from a collision (probably should have hot glued the connector), causing the pipeline to revert back to retroreflective, but that was resolved in update 2024.6.
We used the model provided by limelight which very robust at all of our venues, even with the uneven lighting at DCMP. I’ve also trained my own model in the offseason based on limelight’s guide and a random dataset I found online, and it only took a few hours to have facial detection up and running.
Here is an example on red side where it helped us find a note that was bumped out of position.
987 also used it extensively, which helped recover auto when there were collision on the center line. It was also used to pick up a note dropped at the start of auto.
3 Likes
4476 also used the raw values of the LL3+ with coral in basically the same way as others described.
Took raw tx values, centered the robot by converting field relative joystick inputs to robot centric, and only allowing the robot to move in directions where the intake is centered on the note. We only implemented this for championships, and then more recently stemley cup event where it ran while intaking. Code here.
We also used it similarly in auto for picking up dropped notes/notes that had been nudged/moved. We used it in autos like this.
2056 I believe used LL on color detection, not with coral for higher FPS, though personally I don’t think the FPS was a limiting factor for our use case and we were very satisfied with the performance and would use it again for similar use cases.
We also had reversion issues with pipeline in testing, which luckily never happened during a match, and pit checks made sure the pipeline settings were not reverted.
2 Likes
As in diminished accuracy and precision over distance, seems to be mostly answered by the other posters as well though lol.
we had a fairly low mount angle to assist intakimg right until the note was under the bumper, so we didnt have much distance to see any loss of accuracy over distance. the main issue we had was the latency while turning, had to lock the robot heading while running auto intake
1 Like
We managed to find an incredible model which allowed us to run at the max fps at all resolutions. I truly have no idea how it is this good so if anyone has an idea I’d love to know.
After that you drive robot oriented at tx angle with a PID controller on ty for the speed and a PID controller on tx for the rotation.