Team 254 Presents: FRC 2016 Code

[Tom Bottiglieri]

A quick note on the vision app and the motivation behind using Android.

We started the year under the mindset that we could build a protected zone shooter, but quickly realized with some prototypes and strategy sessions there was serious value to be had by building a small robot that could both go under the bar and shoot from anywhere near the tower. We knew this would require a very good vision system for our robot and got to work trying to make something to run on the NVIDIA Jetson board. This board proved to be very capable of processing frames (in fact, the best performance we got all year was an early prototype running on this board), but had some issues with power up/down reliability. We debated using a computer with a battery built in, but settled on Android because it was cheaper and "cooler".

The app was designed to work well on the hardware we selected for the robot (Nexus 5), but we have seen weird bugs on other devices. For instance, the framerate is worse and the picture is upside down on my Nexus 5X. I'm sure there is a perfectly reasonable cause for this, we just haven't felt the need to fix bugs for platforms that weren't on our robot. If you find bugs in the app or make it work on a new platform, feel free to submit a pull request and our students will review it.

[Eugene Fang]

Originally Posted by Tom Bottiglieri

the picture is upside down on my Nexus 5X.

Apparently the camera module in the 5X is upside down due to packaging reasons. There's a software flag that apps are supposed to read to get camera orientation, but many (like the Augmented Reality feature in eDrawings) don't do it right...

[kingca]

How were you guys able to calculate the traction?

[Jared Russell]

Originally Posted by kingca

How were you guys able to calculate the traction?

I'm not sure what you're referring to? We did have a "traction control" mode that used closed-loop speed feedback along with a gyro to cross defenses while remaining straight, but this didn't require calculating traction.

[wsh32]

How do you check your position in auto after you crossed a defense like the moat, where your wheels might be turning more than you're actually moving? Or did you not run into that problem?

[Jared Russell]

Originally Posted by wsh32

How do you check your position in auto after you crossed a defense like the moat, where your wheels might be turning more than you're actually moving? Or did you not run into that problem?

1) we went slowly enough that worst case slip was limited.
2) we used closed-loop velocity control on the wheels to ensure that even if one side of the drive train momentarily lost traction, we didn't suddenly lose a ton of encoder ticks.
3) in the end, we didn't need to be that precise - our auto-aim could make the shot from anywhere in the courtyard, and on the way back we either just used a conservative distance (for our one-ball mode) or used a reflective sensor to ensure we didn't cross the center tape (for our two-ball modes).

[ranlevinstein]

Thank you guys for sharing your amazing code!
I have a couple of questions:
Why did you choose to follow a path instead of a trajectory during auto this year?
Why did you choose the adaptive pure pursuit controller instead of other controllers?

[Jared Russell]

Originally Posted by ranlevinstein

Why did you choose to follow a path instead of a trajectory during auto this year?

Great question! I assume you are referring to (my) definition of path vs. trajectory from the motion profiling talk (these definitions are hardly universal).

Path: An ordered list of states (where we want to go, and in what order). Paths are speed-independent.

Trajectory: A time-indexed list of states (at each time, where we want to be). Because each state needs to be reached at a certain time, we also know get a desired speed implicitly (or explicitly depending on your representation).

In 2014 and 2015, our controllers followed trajectories. In 2016, our drive followed paths (the controller was free to determine its own speed). Why?

Time-indexed trajectories are planned assuming you have a pretty good model of how your robot will behave while executing the plan. This is useful because (if your model is good), your trajectory contains information about velocity, acceleration, etc., that you can feed to your controllers to help follow it closely. This is also nice because your trajectory always takes the same amount of time to execute. But if you end up really far off of the trajectory, you can end up with weird stuff happening...

With a path only, your controller has more freedom to take a bit of extra time to cross a defense, straighten out the robot after it gets cocked sideways, etc. This helps if you don't have a good model of how your robot is going to move - and a pneumatic wheeled robot climbing over various obstacles is certainly hard to model.

Originally Posted by ranlevinstein

Why did you choose the adaptive pure pursuit controller instead of other controllers?

Simplicity. A pure pursuit controller is basically a P-only controller on cross track error, but somewhat easier to tune. Adaptive pure pursuit is sort of akin to a PD controller (the only difference is a fudge factor in how far you look ahead). If you only have a day to get auto mode working, and the robot is being repaired up on a table while you are coding, then pure pursuit requires very little time to get tuned once you are back on the floor

[apache8080]

Thanks for all of the great resources.

I had a few questions on your vision code:

Are you guys calculating distance from the goal to adjust the hood? If so, how?

If you guys would have used the Jetson TX1, would you have considered using the ZED stereocamera from Stereolabs?

[Jared Russell]

Originally Posted by apache8080

Are you guys calculating distance from the goal to adjust the hood? If so, how?[/url]?

Yep. I'll point you to a few places in the code that help explain how.

First, in the Android app, we find the pixel coordinates corresponding to the center of the goal:
https://github.com/Team254/FRC-2016-...View.java#L131

...and then turn those pixel coordinates into a 3D vector representing the "ray" shooting out of the camera towards the target. The vector has an x component (+x is out towards the goal) that is always set to 1; a y component (+y is to the left in the camera image); and a z component (+z is up). This vector is unit-less, but the ratios between x, y, and z define angles relative to the back of the phone. The math behind how we create this vector is explained here.

The resulting vector is then sent over a network interface to the RoboRIO. The first interesting place it is used is here:
https://github.com/Team254/FRC-2016-...tate.java#L187

In that function, we turn the unit-less 3D vector from the phone into real-world range and bearing. We can measure pitch (angle above the plane of the floor) by using our vector with some simple trig; same thing for yaw (angle left/right). Since we know where the phone is on the robot (from CAD, and from reading the sensors on our turret), we can compensate for the fact that the camera is not mounted level, and the turret may be turned. Finally, we know how tall the goal should be (and how high the camera should be), so we can use more trigonometry to use our pitch and yaw angles to determine distance. We feed these values into a tracker (which smooths out our measurements by averaging recent goal detections that seem to correspond to the same goal).

The final part is to feed our distance measurement (and bearing) into our auto-aiming code. We do this here:
https://github.com/Team254/FRC-2016-...ture.java#L718

Notice that we use a function to convert between distance and hood angle. This function was tuned (many times throughout the season) by putting the robot on the field, shooting a bunch of balls from a bunch of different spots, and manually adjusting hood angle until the shots were optimized for each range. We'd record the angles that worked best, and then interpolate between the two nearest recorded angles for any given distance we want to shoot from.

[Jared Russell]

Originally Posted by apache8080

If you guys would have used the Jetson TX1, would you have considered using the ZED stereocamera from Stereolabs?

Probably not. Stereo was totally unnecessary for estimating range to the goal last year; we were able to estimate our distance to within a few inches using only the method described in the previous post.

[kingca]

oh ok so your "traction control" was making sure that your robot remained straight; by any chance what was the logic behind making the robot stay straight.

[Jared Russell]

Originally Posted by kingca

oh ok so your "traction control" was making sure that your robot remained straight; by any chance what was the logic behind making the robot stay straight.

Yeah - the logic for staying straight is here. There's a PID controller that compares our actual heading to the desired heading, and adjusts the desired velocities of the left and right sides of the drive accordingly.

[frcguy]

Originally Posted by Jared Russell

I'm not sure what you're referring to? We did have a "traction control" mode that used closed-loop speed feedback along with a gyro to cross defenses while remaining straight, but this didn't require calculating traction.

What gyro did you guys use? I saw the Spartan Board on your bot at Chezy Champs, so I assume that you used the ADXRS453 that is built in to it, but I didn't have a chance to get a closer look.

[Jared Russell]

Originally Posted by frcguy

What gyro did you guys use? I saw the Spartan Board on your bot at Chezy Champs, so I assume that you used the ADXRS453 that is built in to it, but I didn't have a chance to get a closer look.

Yes.