Integrating Computer Vision with Motion Control - FIRST Championship Conference

[Jared Russell]

Are you having a tough time coordinating vision with motion?
http://i.imgur.com/8BiHks1.jpg

Does buggy vision make your robot grasp at things that aren't there?
https://static-ssl.businessinsider.com/image/5572060decad04a73f8b4569/drc%20hubo.gif

Do you wish your robot had superhuman powers of perception and control?
http://i.imgur.com/pydaa83.gif

If so, join us at 7:00 PM on Wednesday, April 27 in Room AC265 of the America's Center at the FIRST Championship for our session on "Integrating Computer Vision with Motion Control".

Here's the abstract:

Often the hardest part of solving an FRC computer vision challenge is figuring out how to integrate a camera-based vision algorithm with closed-loop control to automatically point, steer, or drive a mechanism (or the entire robot!). This presentation walks through techniques and best practices that can be employed to mitigate issues like latency, imperfect cameras, and simplified vision algorithms to achieve lightning-fast, precise, and robust control.

This will be presented by Jared and Tom from Team 254, who (together with Austin from Team 971) brought you Motion Planning and Control for FRC (http://www.chiefdelphi.com/forums/showthread.php?p=1474091) last season. We had a great response (and a packed house!) last year, and have incorporated your feedback to make this session even better. To ensure the session is approachable for all levels of experience, we plan on focusing on concrete FRC problems (like auto-aiming a shooter) while showing how concepts like understanding how your camera works, being smart about tuning your vision algorithms, utilizing kinematics, and applying motion control best practices can make your solution better...we want to help teams compete while also opening your students' eyes to how professional roboticists approach similar problems.

And the great news is that all Championship Conference talks will be recorded this year, so you don't need to choose between this session and Karthik's...you can have both!

[D.Allred]

Thanks Jared. Hopefully I'll be able to attend. Last year's motion profile course was terrific. We were able to incorporate the concepts in an off season project and this year's robot. Look forward to learning more about vision control.

David

[generaldgibson]

Sounds cool, I'm fairly new to programming but hoping to learn about vision tracking. Can't wait!

[orangemoore]

Does anyone have a time machine?

http://www.chiefdelphi.com/forums/showthread.php?t=147028

Otherwise this is going to be a tough choice.

Will be recorded again if I don't make it to this session?

[generaldgibson]

Does anyone have a time machine?

http://www.chiefdelphi.com/forums/showthread.php?t=147028

Otherwise this is going to be a tough choice.

Will be recorded again if I don't make it to this session?

As OP said, the sessions will be recorded this year - I'm planning on watching one of the two as well :p

[orangemoore]

As OP said, the sessions will be recorded this year - I'm planning on watching one of the two as well :p

I should really read better. ::ouch:: ::rtm::

[GreyingJay]

And the great news is that all Championship Conference talks will be recorded this year, so you don't need to choose between this session and Karthik's...you can have both!

YES! This is awesome news. Will they be put up on YouTube or something similar?

There were a whole lot of great looking conferences and since our team didn't make it to Champs I was very tempted to fly out anyway just to experience the sessions.

[Tom Bottiglieri]

If anyone has any particular questions on this topic, feel free to post them here. We think we have a good idea of the problems most FRC teams run in to, but it's always good to learn a bit more. We will do our best to address these topics in our presentation!

[rich2202]

Our main problem is how long it takes to process an image. By the time you process the image. The robot is somewhere else.

Thanks for recording. I'll be robot inspecting, and will miss the Wednesday conferences.

[Jared Russell]

Our main problem is how long it takes to process an image. By the time you process the image. The robot is somewhere else.\

We will DEFINITELY deal with this one :)

[Ether]

And the great news is that all Championship Conference talks will be recorded this year, so you don't need to choose between this session and Karthik's...you can have both!

That is great news.

Do you know if the speaker will be fitted with an on-person mic that feeds electronically into the recorder? Audio quality from acoustic pickup from an on-camera mic at the back of the room is close to inaudible for those of us with aging ears.

[alex.lew]

Do you know if the speaker will be fitted with an on-person mic that feeds electronically into the recorder? Audio quality from acoustic pickup from an on-camera mic at the back of the room is close to inaudible for those of us with aging ears.

young ears too :)

the recordings will be a nice treat post-finals week. thank you for putting this together!

[jreneew2]

One question that I hope you will cover is:

When to use vision and when to not use vision. Last year, we had a mentor work on tracking the yellow totes for a couple of weeks and we didn't end up using it at all. :/

However, this year you need vision tracking if you have any hope of shooting a high goal in autonomous.

Thanks,
Drew

[marshall]

One question that I hope you will cover is:

When to use vision and when to not use vision. Last year, we had a mentor work on tracking the yellow totes for a couple of weeks and we didn't end up using it at all. :/

However, this year you need vision tracking if you have any hope of shooting a high goal in autonomous.

Thanks,
Drew

That's a strategy problem. You have to choose to focus on a task first and then you move to how to use vision to augment or automate performing that task.

I am looking forward to this talk though. I don't know what questions I'll have yet but I know that I'm going to have some.

[FlamingSpork]

However, this year you need vision tracking if you have any hope of shooting a high goal in autonomous.


My team's robot doesn't have a functioning vision system, and yet still fires a high goal shot in auto. We used a hard-coded position to aim our turret from the spy box.

[Jared Russell]

Do you know if the speaker will be fitted with an on-person mic that feeds electronically into the recorder?

Not sure, but I'll bring this up.


When to use vision and when to not use vision. Last year, we had a mentor work on tracking the yellow totes for a couple of weeks and we didn't end up using it at all. :/

While this is not a strategy talk first and foremost, we will most certainly discuss this. In particular, we will look back at the history of vision in FRC and you will quickly notice how often the vision challenge is a diversion for the majority of teams.

[JesseK]

Mostly likely, I won't be able to attend the conference (either session at 7pm), but I do have a question and a few examples of what I mean.

What are some rules of thumb for the prerequisites to attempting vision?
- Great PID Control (fast settle, little/no overshoot) or "Good Enough" PID Control
- Level of precision on drive train turning or (auton) "drive straight" distance offsets
- For high-load situations, is a braking mechanism necessary

[dcarr]

While this is not a strategy talk first and foremost, we will most certainly discuss this. In particular, we will look back at the history of vision in FRC and you will quickly notice how often the vision challenge is a diversion for the majority of teams.

I think it's interesting that Stronghold seemed to present the "perfect storm" of factors that made vision a viable and extremely value asset for a larger-than-ever subset of teams.

Looking forward to this talk!

[Tom Bottiglieri]

I thought I'd share a clip we took today prepping for this conference presentation. We will definitely be going through all the pieces it takes to get a robot tracking like this!

http://i.giphy.com/l3V0suUzaCDToQK6k.gif

[billbo911]

Tom and Jared, will you guys be discussing the realities of being able to process frames faster than the camera can produce them?

For example, a camera's maximum frame rate is listed at 30 frames per second, but your vision processing board can process over 80 frames per second.

[James Tonthat]

Too cool for school!

::ouch::

[marshall]

I thought I'd share a clip we took today prepping for this conference presentation. We will definitely be going through all the pieces it takes to get a robot tracking like this!

http://i.giphy.com/l3V0suUzaCDToQK6k.gif

I didn't know you guys had ILM as a sponsor. ;) I kid.

Seriously cool and I'm really looking forward to this talk now.

[billbo911]

Some additional questions.
What is the minimum frame rate you feel is acceptable for FRC?
How much lag is acceptable and how do you cope with it.
What is the most efficient way to get targeting data into the RoboRio (TCP, UDP, Serial etc.)?

[pmangels17]

I think it will be addressed, but I would be curious to hear exactly how much control on 254 is done by drivers vs with vision. What are your failsafes if the vision board suddenly crashes/shorts/explodes?

Also, what other sensors do you combine with vision for a more complete picture of your control system. For example, if you want to know distance to a target, are you using only a camera for that, or are there other sensors that you use to crosscheck your measurements.

And, specifically on your robot, how do you relay to the drivers that everything is in position for a shot? Is there a giant green check mark, does the system just not let you fire until it's ready, is the firing also controlled autonomously, or do you have a different solution?

[Jared Russell]

I think it will be addressed, but I would be curious to hear exactly how much control on 254 is done by drivers vs with vision. What are your failsafes if the vision board suddenly crashes/shorts/explodes?


The operator holds a button saying they want to auto aim, and the driver holds a button saying that it's okay to shoot. Everything else is autonomous.

[JesseK]

The operator holds a button saying they want to auto aim, and the driver holds a button saying that it's okay to shoot. Everything else is autonomous.

Hmm - including sensing on the intake and ball pre-positioning?

[Jared Russell]

Hmm - including sensing on the intake and ball pre-positioning?

The intake also has some automation but no, I was speaking about the aiming and shooting subsystem.

[wesleyac]

Do you guys find that it's easier to accurately position the turret than it is to turn the drivetrain? That gif is really impressive, and I'm wondering how much of that is the turret being mechanically easier to control. Do you do any motion profiling on it, or is it just PID?

I'm looking forward to your talk!

[Jared Russell]

Do you guys find that it's easier to accurately position the turret than it is to turn the drivetrain? That gif is really impressive, and I'm wondering how much of that is the turret being mechanically easier to control. Do you do any motion profiling on it, or is it just PID?

I'm looking forward to your talk!

The turret is just running position PID (onboard a CAN Talon SRX). Turrets are much easier to control than a drivetrain (inertia, friction, and skidding are basically non-issues).

[billbo911]

The turret is just running position PID (onboard a CAN Talon SRX). Turrets are much easier to control than a drivetrain (inertia, friction, and skidding are basically non-issues).
I assume you are using the camera to generate the set-point. Are you closing the loop with the camera as well, or are you using some other sensor like a gyro or encoder?

[4AngryNutjobs]

Cannot wait for the talk! My team has struggled inplementing vision in the past, and I am sure this conference will help a bunch! Once the video is posted can someone share the link below?

[Cog]

Are you having a tough time coordinating vision with motion?
http://i.imgur.com/8BiHks1.jpg

Does buggy vision make your robot grasp at things that aren't there?
https://static-ssl.businessinsider.com/image/5572060decad04a73f8b4569/drc%20hubo.gif

Do you wish your robot had superhuman powers of perception and control?
http://i.imgur.com/pydaa83.gif

If so, join us at 7:00 PM on Wednesday, April 27 in Room AC265 of the America's Center at the FIRST Championship for our session on "Integrating Computer Vision with Motion Control".

Here's the abstract:


This will be presented by Jared and Tom from Team 254, who (together with Austin from Team 971) brought you Motion Planning and Control for FRC (http://www.chiefdelphi.com/forums/showthread.php?p=1474091) last season. We had a great response (and a packed house!) last year, and have incorporated your feedback to make this session even better. To ensure the session is approachable for all levels of experience, we plan on focusing on concrete FRC problems (like auto-aiming a shooter) while showing how concepts like understanding how your camera works, being smart about tuning your vision algorithms, utilizing kinematics, and applying motion control best practices can make your solution better...we want to help teams compete while also opening your students' eyes to how professional roboticists approach similar problems.

And the great news is that all Championship Conference talks will be recorded this year, so you don't need to choose between this session and Karthik's...you can have both!
I'm sorry if I missed it, but do you know where will the recordings of these seminars be made available?

[axton900]

^ Where will the recordings be located???

[Steven Smith]

Last year's might have been posted elsewhere, but it was on Youtube (https://www.youtube.com/watch?v=8319J1BEHwM)

Let's keep in mind that these presenters are on top tier teams that likely need their support for the next couple days as they compete. Last year's video was posted ~5-6 days after champs ended, and while I can't speak for the presenters, I would assume you can expect it sometime in the next week or so with a link to it in this thread. I wouldn't want to try to upload that large of a video via hotel internet myself.

That being said, if you haven't watched last year's or if it has been awhile, you can pass the time with the 2015 video which was also very awesome and applicable. I'd argue for the majority of the teams, maybe more applicable than this year's.

[Ether]

And the great news is that all Championship Conference talks will be recorded this year, so you don't need to choose between this session and Karthik's...you can have both!

Someone said they were short so only recording one per timeslot
...

[BrianAtlanta]

Sadly, I wasn't able to attend the session since I had to fly out Thur morning. When a video is available, if someone could post a link, I would appreciate it.

Thanks,
Brian

[Jared Russell]

One way or another I will make sure a recording of this session (of sufficiently high quality) gets posted online. We will just redo it from home (and answer the questions that we received in person) if necessary.

[Ether]

One way or another I will make sure a recording of this session (of sufficiently high quality) gets posted online. We will just redo it from home (and answer the questions that we received in person) if necessary.

Thank you. Very much appreciated.

[Jared Russell]

Slides are here: https://goo.gl/mppi4E

[orangelight]

Slides are here: https://goo.gl/mppi4E

Thanks

[orangemoore]

It says I don't have access

I must be special because I can get into it.

Also it is an amazing presentation.

[Ozuru]

One way or another I will make sure a recording of this session (of sufficiently high quality) gets posted online. We will just redo it from home (and answer the questions that we received in person) if necessary.

Thank you very much!

[dcarr]

Here's a phone recording I took of the talk - I missed a minute or so of the intro, and a few questions at the end, but the rest is here.

https://youtu.be/rLwOkAJqImo

Jared and Tom, thanks for a great talk!

[billbo911]

Slides are here: https://goo.gl/mppi4E

This is FANTASTIC! Thank you!

[Adithya Balaji]

Here's a phone recording I took of the talk - I missed a minute or so of the intro, and a few questions at the end, but the rest is here.

https://youtu.be/rLwOkAJqImo

Jared and Tom, thanks for a great talk!

This is awesome!! Thank you for posting this.

[Pault]

Here's a phone recording I took of the talk - I missed a minute or so of the intro, and a few questions at the end, but the rest is here.

https://youtu.be/rLwOkAJqImo

Jared and Tom, thanks for a great talk!

Audio on this is perfectly fine, and paired with the slides on a second monitor I was able to get everything out of the presentation.

This was some pretty awesome stuff. I was hoping to have the team work on a new co-processor based vision system over the offseason, and I can see this presentation influencing a lot of it. There was really something in that presentation for everybody; they managed to cover the basics while still focusing on things like predictive vision and how to actually determine the real world coordinates of the goal. Another home run talk by the cheesy poofs; I can't wait to see what they come out with next year.

[virtuald]

I flew in late on Wednesday, so I'm sad I missed this and I haven't heard the audio yet. However, the presentation slides are quite excellent!

I know that in 254's current tracker you're using an external processor and aiming for maximum performance for your image processing -- but I'm curious if you've tested the current hardware with a lesser performing camera setup. For example, on the RoboRIO I estimate our image processing code is 320x240 @ 10fps with ~100ms latency (haven't measured lately, but I seem to recall that's what I got when I did measure). How much does your turret tracking degrade with that kind of lower performance? Or does the latency compensation and high quality construction of the turret compensate for that?

[Jared Russell]

How much does your turret tracking degrade with that kind of lower performance? Or does the latency compensation and high quality construction of the turret compensate for that?

It would work fine. The fact that the goal is static, and you can estimate your robot's movement VERY accurately over a few hundred milliseconds, combine to mean that you can do really well even with lots of latency and limited resolution.

[Cabey4]

Has anyone found the magical archive of the talks? I was really looking forward to getting to see this one and karthik's :(
Luckily, you guys have promised getting it online, but if someone could have a link to all of them that would be amazing. There dosen't seem to be anything on the FIRST champs website.