@ gr0p3r

Thanks! I made the project that aleph8nought mentioned.
My light also fades from color to color, I had to up the pwm base frequency after dropping the clock to 1MHz.
The code is linked in the blog post for specifics...

Or the tsop58038 which is the replacement for the tsop4038...
I have used both for bootloading and lacking AGC they don't turn off with extended duty cycle..

IR bootloader: http://n0m1.com/2012/05/07/superduplex-an-infrared-bootloader-for-arduino/

Nice job

Thanks! 

It was quite a bit of work getting everything to play nice together, especially the half duplex timings and the phase delay...
Looks like it just made Hackaday: http://hackaday.com/2012/05/07/program-your-arduino-via-ir-using-the-arduino-ide/

I've just finished up the post for the IR bootloader.
The maximum reliable speed obtained is around 4000 baud, not fast but not terrible...

The info, source and video is here:
http://n0m1.com/2012/05/07/superduplex-an-infrared-bootloader-for-arduino/

Is the entire schematic available also?

Yes, I just added to the blog post: http://n0m1.com/2012/04/30/how-to-bread-board-arduino/

I've posted my eagle library here, which has a physical layout corresponding to the chip.
It might be useful: http://n0m1.com/2012/04/30/how-to-bread-board-arduino/

So, do I get the TL072 input from the + of the thermistor (i.e. the field side of the thermistor), or from the MCU side which will be after the resistor (i.e. the midpoint)?

You want to measure at the midpoint as that is the output of the divider.
If you measure without the divider, you will be measuring the supply voltage on the other side of the thermistor, not what you want...

Take a look at the picture at the top of this post:
http://n0m1.com/2012/01/28/adc-battery-voltage-divide-match-and-measure/
This is for measuring battery voltage, but in your case R1 or R2 will be a thermistor and the voltage will vary with temperature. 
The post explains the need for the capacitor and the math to calculate the output voltage as well... 

Does having a thermistor on all the time use a lot of current?

Depends on the resistance of the thermistor and the voltage passing thru it.
You will have the thermistor in a voltage divider: http://en.wikipedia.org/wiki/Voltage_divider#Resistive_divider
So add the resistances in series and calculate power with P = V*V/R.

You should be able to share the thermistor but I'd use a cap to ground from the thermistor/divider output to reduce input impedance.

Use a resistive divider to reduce it. http://en.wikipedia.org/wiki/Voltage_divider#Resistive_divider

Connect your gnds so that you have a common reference and use large value resistors so that you don't dissipate a lot of power and burn up the resistors.
10k and 47k would be a good choice but you should make sure the input is really 5.6 volts as it could be higher when it's unloaded ...

is there a way to prevent power dissipation?

Yes, use a logic level N-channel MOSFET to switch the low side. You want logic level as it won't require more than 5v to turn on fully.
A FET with a low Rds(on) will dissipate much less that a transistor, you can calculate dissipation with P = V*V/R.

If you want SMD, the IRLML2402 is a nice part in SOT23. http://www.irf.com/product-info/datasheets/data/irlml2402.pdf

For through hole parts, Digikey's parametric search comes up with many choices, the first one is good for 600mA at 5v gate, second one is quite beefy with a really low Rds(on):
http://www.digikey.ca/product-detail/en/ZVN4206A/ZVN4206A-ND/92604 
http://www.digikey.ca/product-detail/en/NTD4960N-1G/NTD4960N-1GOS-ND/2409565

Here's a more conventional example:
http://letsmakerobots.com/node/865

So I just take the x,y,z values from here:
http://code.bildr.org/download/959.zip
and plug them into here directly?  Combining the 2 sketches.

Yes, that is correct. Raw values or scaled doesn't matter, raw is faster and more accurate.
In process, the library scales everything up with an equal multiplier to reduce quantization losses.

sixDOF.compCompass(-(raw.ZAxis), -(raw.XAxis), raw.YAxis, -(accel.z()), -(accel.x()), accel.y(), true);
Mine are raw values also?
Why do the first 2 - Z,X have minus signs, Y does not?
Do I include these signs before my x,y,z too?

You may want to play with the arrangement and sign of xyz depending on the orientation of your accelerometer and compass.
My PCB is mounted vertically and the xy of the accelerometer was not the same as the xy of the compass.
This is why I have (-z, -x, +y).  The default order is (+x, +y, +z).

I don't see how this can work for small tilt angles, because the Z value doesn't change, in my case.
Can you explain how it works without this value?
Let's say you have tilt and X,Y from HMC5883, but no Z.

The libraries trigonometry is non-linear and ratiometric.
Non-linear as in on axis effect is the lowest, while a right angle axis will have the highest effect.
Ratiometric as in a change in only one axis effects the ratio and therefore the end result.

Does it look like this?:
http://www.ebay.com/itm/FAULHABER-DC-12V-Motor-2342L012CR-Gear-64-1-Encoder-12CPR-free-ship-/130585677299

The 12CPR likely indicates it's using a pair of hall sensors and a multipole disk magnet for the encoder.
Which would be 64:1 x 12 CPR = 768 CPR at the shaft. With the gear head that is a pretty good count.

Edit: Just saw the back of the motor, it's slotted disk and opto, not multipole magnetic and hall...

The main points of your code look good.
You are using interrupts to read the encoders.
The PID loop is on constant timebase.

My approach would be to run PID on each motor individually, which also makes for an incremental approach.
And command them to turn by specifying a different number of steps for each wheel for the same time.

If the behavior is to run away, try reversing either the motor wires, or reversing the direction bit in sw...
What is the transition count of your encoders per revolution? Low count can be tricky...

I don't believe the IR sensors are simple photodiodes, they communicate with the door opener with some proprietary protocol.

If this is the case, you might try making your own.
Modulate an IR led at 38khz, use a lens to extend range...

And receive it with either one of these demodualtors, TSOP4038 or TSOP58038.
These two specific parts are rated for continuous use, I found both at Mouser.
Regular TV remote demodulators will stop responding to a continuous signal (over 20% duty cycle) after a short while due to the gain control.