The Perfect swerve

[Gdeaver]

In 2013 looked at high end gaming mice. With swerve and the chassis orientation decoupled, there are 2 solutions of the x and y counts coming from the mouse. On competition carpet and no changes to the optics we found the accuracy to be less than needed. The optic flow algorithm is not tuned for this use. With a usb port on the roborio in 2015, A usb camera highly filtered and a more robust optic flow algorithm may yield better results. A gyro at minimum would need to be fused with the optic flow. This is all for field centric control. There are 2 paths 2 look at. Sensing from the robot reference frame (gyro accelerometer fusion). Adding a world reference frame with a magnetometer or some other sensor to reference out side the robot frame of reference. GPS is out. Constellation navigation has grabbed my curiosity. It's hard. In 2013 we could have reset the gyro every time we went up against the feeder station wall to correct for drift. This year we considered IMU field centric control not doable because of the constant impacts and never having time for a reset. The last thing our drivers need this year is for the field centric control to suddenly shift several degrees while being smash defended and trying to roll out. With our low designs the last several years a magnetometer location and calibration issues ruled out that solution. Fortunately for the future of swerve field centric a couple of companies have released affordable IMUs based on gyro, accelerometer and magnetometer sensors coupled with highly tuned extended state kalman filters that should handle the rough First environment. The key is constant hard and soft iron calibrations on the magnetometer sensor. I'm hoping to make this an off season project if I can get some programming students on board. I think soon a plug and play Field centric IMU solution will be available for FIRST.

[Ether]

Originally Posted by Gdeaver

there are 2 solutions of the x and y counts coming from the mouse

It's not clear what you mean by 2 solutions. With a single mouse, an XY displacement could correspond to a multitude of possible robot motions.

With 2 mice, you could in theory derive all three degrees of freedom of the robot motion, if the XY readings are accurate.

[ekapalka]

It seems to me like follower wheels shouldn't be out of the question when monitoring position, especially when used with other forms of sensory. Here is how you calculate the speeds, and this is a thread all about how to utilize and manage that information (both thanks to Ether).

[Gdeaver]

There have been several college papers written about this subject and the ones that had some success used 2 mice and a gyro. There are some crowd sourced devices on the market based on these. They work kind of on small robots and smooth surfaces going slow. Haven't seen anybody that has found a mouse solution that would work in the First environment.

[Gdeaver]

The problem with First is high G impacts.

[RyanCahoon]

Originally Posted by Gdeaver

A usb camera highly filtered and a more robust optic flow algorithm may yield better results

Without specialized hardware and/or a lot of custom optimization, I'm doubtful that you'll be able to run an optical flow algorithm much faster than 15 Hz at any decent resolution. At that framerate with a robot traveling 12 ft/s, you'll see displacements of 10 inches/frame. Getting an unobstructed view of that much carpet beneath your robot seems like it would be a challenge, assuming you can even reliably track displacements that large using the texture of the carpet.

[s_forbes]

Originally Posted by RyanCahoon

Without specialized hardware and/or a lot of custom optimization, I'm doubtful that you'll be able to run an optical flow algorithm much faster than 15 Hz at any decent resolution. At that framerate with a robot traveling 12 ft/s, you'll see displacements of 10 inches/frame. Getting an unobstructed view of that much carpet beneath your robot seems like it would be a challenge, assuming you can even reliably track displacements that large using the texture of the carpet.

As an alternative method: what's the average height of a ceiling at a typical venue? The truss structures I've seen in most gymnasiums might impact an optical flow approach, but it may work better than the carpet.

[ekapalka]

Originally Posted by s_forbes

As an alternative method: what's the average height of a ceiling at a typical venue? The truss structures I've seen in most gymnasiums might impact an optical flow approach, but it may work better than the carpet.

Astral navigation :D I really wanted to use this approach to verify the gyro readings when in view of landmarks on the ceiling, and use the gyro for orientation when the landmarks are obscured (by the truss / balls flying overhead), but we never got around to getting a camera for it (not even a stereo camera - just an Axis). If anyone ever makes any progress on this, I would love to know about it. I've only ever seen one real-world thing that boasts their use of it, and those are StarGazer localization cameras, which need specific landmarks (so this particular method of astral navigation isn't suited to be used on the playing field)

[Gdeaver]

What I was thinking of is to use the lights as constellations. You don't need color or features, just a filtered matrix of 1 and zero's.

[RyanCahoon]

Originally Posted by Gdeaver

What I was thinking of is to use the lights as constellations. You don't need color or features, just a filtered matrix of 1 and zero's.

Integrating this with gyro readings using a Kalman filter would be a fun project.

I've thought before (it may have been mentioned before on CD) about trying to track field position based off of the driver station lights: look for 3 red/blue lights with predefined spacing between them.

[Gdeaver]

This year several teams have developed swerve that works very well. Right now I think the biggest problem now is not the physical swerve but the human interface and the driver. Giving the driver an extra degree of freedom really loads up the drivers brain when they are in a match and performing in a match. We over come the the human problem part of the system by going to as many off season events as we can to train our drivers. Going forward with swerve improvements needs to focus on the human part of the equation. Rock solid field centric control would be one path. A swerve simulator was posted on CD. It's primitive. A real good simulator that captures the physics of swerve could be used to help train drivers. That's what the military does. Right now I believe that a reliable IMU based on inexpensive MEMs sensors is almost there. This summer I think we will look at the IMU' and algorithms that just came to market to see if they will work.