Imagine for one second that you are a UAV developer. The DIYdrones ArduPilot is an awesome piece of hardware; you love all the great sensors that are attached; you love the all-in-one design and the small/light size.  But you are also feeling the limits of it’s ATMega 2560 processor and the limits of it’s 256Kb of RAM.  And maybe, you enjoy developing code within a full blown Linux developers environment with all the supporting libraries and device drivers that Linux has to offer.

Hardware Layout

Here is a picture of a protype “hybrid” autopilot system for small UAV’s.  What you see in the picture includes:

• An ArduPilot Mega 2.0 (lower left)  The APM2 includes an MPU-6000 IMU, a Mediatek 5hz GPS, static pressure sensor, compass, and a variety of other items.
• A 1Ghz Gumstix Overo with a Pinto-TH expansion board. (longer thinner board, upper right).  The Overo runs Linux, has 512Mb of RAM, and has a hardware floating point processor.
• An MPXV7002DP pressure sensor (upper left).  This is attached to a pitot tube and measures airspeed.
• A Sparkfun TTL level translator (between the overo and the pressure sensor)
• A Futaba FASST 6 channel receiver (lower right)
• A set of power distribution rails (8×3 block of 0.1″ pins.)
• [not shown/mounted remotely] Digi Xtend 900Mhz radio modem.

The above picture was deliberately taken without most of the wires to show the components.  Once all the wires are connected, things will look much more “busy”.  This is a prototype system so the components will be jumpered together, probably with a bit of hot glue here and there to ultimately secure the wiring before flight.

Software Layout

Just as important as the hardware layout is the software layout.  This is maybe a bit more boring and I can’t really take a picture of it to post here, but I will give a few bullet points for anyone who might be interested.

• The APM2 runs a custom built firmware that simply collects all the sensor data, packages it up into checksummed binary packets, and sends the data out over a uart.  In addition, the APM2 reads checksummed binary packets inbound over the same uart.  Generally inbound packets are actuator commands.
• The Overo runs a custom built autopilot application that is portable.  It is not device/hardware independent, but it is designed to talk to a variety of hardware over a common interfaces.  This Linux-based autopilot app can read all the sensor data from the APM2 (being transmitted at 100hz over a 500,000 baud link.)  The autopilot app runs a computationally intensive 15-state kalman filter.  Manages the mission tasks, computes wgs-84 great circle navigation targets, runs the low level PID code, and ultimately sends servo position commands back to the APM2 over the same 500,000 baud link.
• The high level autopilot software runs in an environment that is more natural and has more power/space for developing the “high concept” portions of the code.  The APM2 is great for reading a bunch of sensors really quickly.  In a hybrid system both major components are able to do what they do best without having to suffer through tasks they aren’t as good at.
• The high level autopilot is configurable through a structured XML configuration system.  It includes an advanced 15-state kalman filter.  It runs the same PID code configured in the same way as the FlightGear flight simulator.  It also includes the same “property system” that FlightGear pioneered.  It includes a rich set of IO API’s, and hopefully will soon sport a built in scripting engine.
• A word about the “property system”:  The property system is an in-memory tree structure that relates “ascii” variable names to corresponding values.  There is loose type checking and on-the-fly type conversion similar to many scripting systems (I can write a value as a ‘string’ and read it back out as a ‘double’ and the system does sensible conversions for me.)  Any internal module can read/write the property tree, so it is a great way to build “loose” interfaces between modules.  Modules simply reference the values they need and there is no compile time enforcement.  This allows for building “robust” interfaces and eliminates much of the cascading changes throughout the code when a class gets updated.  In addition, we can build external network interfaces to the property system which allows external scripts or other code to monitor and even modify the property tree.  This allows external scripts access to the core values of the sim and access to set modes and otherwise “drive” what the autopilot does.  In addition, there is a natural one-to-one mapping between the property tree and an xml config file.  This is handy for easily loading/saving configurations.  At a low level, property system variable access from the core code is simply a pointer dereference, so using these values is almost as fast as referencing a locally scoped variable.  It is hard to describe in a few words what all the “property system” is or does, but it’s an amazing infrastructure tool that helps build flexible, robust, and high functioning code.

September 7, 2012 Update:

Here is a picture showing the whole system plugged together with 6″ sparkfun jumper wires.  The wiring gets a little “busy” when everything is connected together so I plan to do some work buttoning things up, simplifying, and cleaning up the wire runs.  I’ve attached 3 servos to test auto vs. manual outputs.  You can see a 900Mhz Xtend modem (1 watt power, 115,200 baud), and the entire system is powered by a single RC receiver battery.  In the aircraft this should all run just fine off the BEC.  And here is a video of everything in action:

That’s about it. The modified APM2 firmware is happy and running.  The overo autopilot code is up and running and talking to the APM2.  Next up is validating the radio modem ground station connection, and then I’m running out of things to do before dropping this into an airplane!

September 17, 2012 Update:

It’s far from a perfect integration, but I spent some time this evening bundling up wires.  Definitely an improvement.  Working towards full integration into a Senior Telemaster (electric powered) and hope to do some first test flights later this week.  Today I test fit everything, powered everything from the main flight battery and verified basic control surface movements in manual mode and autopilot mode.  The radio modem link is running and the mapping and instrument panel views on the ground link up and run.

Here are a few random notes on the APM2 and servos.

As we all know, the servos are controlled by sending a digital pulse on the signal line to the servo.  The length (time) of the pulse maps to the position of the servo.  A 1500us pulse is roughly the center point.  900us is roughly one extreme and 2100us is roughly the other extreme.  Different systems will use slightly different numbers and ranges, but these are good numbers to start with.  Historically, RC systems stacked the pulses for multiple channels on a single radio signal — affectionately called the “pulse train.”  If you crunch the numbers — given a max pulse width of about 2000us and a bit of inter-pulse spacing, we can send about 8-10 channels of data at 50hz on a single signal line.  Standard analog servos are designed to expect 50 pulses per second and can misbehave if we send a faster signal.

With modern 2.4Ghz systems, the servo/channel data is sent digitally and at much faster data rates.  Receivers drive each servo pulse concurrently and the concept of a “pulse train” is largely gone — as is the need to space things out at a 50hz rate.  If you crunch the numbers again and look at a max pulse width of 2000us with a bit of inter-pulse space, we can actually send servo data at 400hz to a digital servo and it will respond as expected.  (8 channels at 50hz, 1 channel at 400hz — makes sense.)

What does all this mean to an UAS builder?

• It means that if we use digital servos, we can send position updates to the servo and expect some response at up to a 400hz rate. In real life these servos have mass, motors that take time to move, on board electronics that take time to process the signal, and other factors, so we can’t expect instantaneous response, but we can expect an improvement over the old 50hz analog servos.
• It means we have a strong motivation to sample our sensors faster than 50hz, run our attitude filter faster than 50hz, run our guidance, control, and navigation code faster than 50hz, and command our actuators faster than 50hz.
• It means that if we can increase our throughput to 100hz or 200hz, we can have a far more responsive and “tight” system than we could if everything was running at 50hz.  Faster response to gusts and disturbances, ability to counter more of the aircraft’s natural dynamics.  With proper autopilot tuning, this can lead to a much higher quality system, much more accurate flying, and a much more stable platform to hang our sensors off of.

The APM2 has the ability to set the PWM frequency for the output channels.  Channels are grouped and several hang from a single clock generator so we can’t change the PWM rate individually, but we can do it for groups of channels.  Question: if our system is running at a full 100hz update rate, does it make sense to set the servo output rate to 100hz or something higher?  I argue that we should set the output rate as high as is possible (400hz) because the servo will pick up a change fractionally quicker at a 400hz output rate than at a 100hz output rate.  In the end though, most servos are somewhat noisy and crude and subject to physics, so it may not have a huge noticeable effect, but any little thing that makes the system more responsive and increases the end to end throughput from sensing to responding is welcome.

The autopilot architecture I am building involves and APM2 collecting all the sensor data and sending it over a serial connection to a Gumstix Overo running Linux, and then in return the Overo sends servo commands back to the APM2.  As you can guess, this turns out to be quite a bit of data being sent at a high rate.  If all the processing, filtering, and computation is being done on the Overo it is important to have a high update rate, low latency, reliable communication, and no major pauses in processing at either end.

In this architecture, the APM2 isn’t running the DCM filter.  It is not logging to the SD card, it is not communicating with the ground station.  It is simply collecting all the sensor data and blasting it across the data pipe to the Overo.

First let’s do a little math.  I have 6 packet types (so far) of varying sizes: Pilot Input = 16 bytes, IMU = 14 bytes, GPS = 36 bytes, Barometer = 12 bytes, Analog Inputs = 10 bytes, Actuator Commands = 16 bytes.  In addition, each binary packet has 2 start bytes, a packet ID, a size byte, and a 2-byte checksum — a six byte wrapper.  Assuming we send one of each packet, that would be 140 bytes.  Assuming we want to send a set of these packets 100 times a second, that is 14,000 bytes per second.  (I realize we don’t send the gps data at 100hz, but for the sake of simplicity and margins, let’s pretend we do.)  Now assume for the purposes of computing required baud rates, there are 10 bits per byte.  That means we would need to run at least at 140,000 baud to handle all the traffic we need.

Many software packages (including the arduino dev system and the arduino docs) seem to think that 115,200 baud is that maximum rate possible, so this could be a problem!

The next “standard” baud rate above 115,200 is 230,400.  It turns out that the APM2 FastSerial code happily let’s me specify this baud rate and my Linux based Overo does as well.  I tested this with the USB console connection and it works.  (Or so I thought.)

My next task was to get the communication link running on a direct UART connection.  This will simplify aircraft integration, reduce cable weight, etc.  However, as soon as I moved to a direct connection, the APM2 could no longer see data coming back from the Overo at 230,000 baud.  What is going on here?

The standard baud rates since the beginning of time have always been 300, 1200, 2400, 4800, 9600, 19200, 38400, 57600, 115,200, 230,400, etc.  However, when I dug into the Arduino FastSerial code and the Mega 2560 data sheet I discovered that the Arduino does things completely differently.  Baud rate is defined relative the system clock divided by an integer value.  When you ask for 115,200 baud, FastSerial crunches that value through a funny formula, comes up with an approximate integer divider and sets the baud based on that.  So when you crunch the math, what baud rates does the Arduino actually support?  It turns out you can do: 1,000,000, 500,000, 333,333, 250,000, 200,000, 166,666, 142,857, 125,000, 111,111, 100,000, 90,909, … etc.

So here’s the funny thing.  The uarts at both ends have some flexibility and can successfully communicate even if the other end isn’t exactly on the correct pace.  When you request 115,200 baud on the APM2, you are actually getting 111,111 baud!  But the other end can usually handle a little slop and no one ever notices.  The problem is that the arduino cannot do 230,400 baud — the options are either 200,000 or 250,000 baud and both of them are too far away from 230,400 to handle the timing slop.  In my case, the Overo actually could handle the APM2 output, but the APM2 couldn’t read the 230,400 baud incoming data from the Overo.

!@#$@#$!@!!!!

The next question I asked myself was how hard would it be to rewrite the linux serial drivers to support different baud rates, but quickly I decided that was just a path I didn’t want to head down unless there was absolutely no other option.

Finally I started looking to see if there were any matching (or very close) baud rate options faster than 115,200.  The only one I could find was 500,000 baud.  But that’s crazy fast and couldn’t possibly work, right?  Well if my other option is to start hacking linux serial drivers, I decided I could at least set the baud on both ends to 500,000 and see what happens.

And the answer is:  IT WORKED!

So I’m running 500,000 baud between the APM2 and the Gumstix Overo.

There are questions though — can the APM2 keep up with this?  Can the Overo keep up with this? What happens if the system is busy processing interrupts.  Are the ring buffers big enough to keep up?  Will we drop data?  Are there other hidden places where this could all break down or blow up?  I think the answers to those questions remain to be worked out and depend quite a bit on the loads at each end of the pipe.  This gets back to the delicate balance required by embedded systems and the requirement to carefully work your way forward and understand the systems at all levels in order to achieve a solid result.

Question: if 100hz updates require about 140,000 baud, then 200hz updates would require 280,000 baud — hmmm.  Does this mean it would be possible to sample the sensors at 200hz, send the data to the 1Ghz Overo at 200hz, run my kalman filter attitude estimate at 200hz, run my autopilot PID’s at 200hz, my navigation code at 200hz, and send actuator commands back at 200hz?  (Assuming digital servos that can respond to PWM rates as high as 400hz.)

Update: September 4, 2012

I have reconfirmed that the Gumstix Overo cannot transmit two serial streams simultaneously.  I use one serial port for the APM2 interface and one serial port for the wireless ground station interface.  The issue is that if I dump a big chunk of data @ 115,200 baud from the flight computer to the ground station, no APM2 communication can happen until that transaction flushes through.  The strategy I will employ for now is to stuff ground station messages into a local (user space) buffer and then trickle a few bytes out per frame to the serial driver.  This forces me to pace the ground station messages so I don’t overflow my application level buffer (which would reject new messages until there is sufficient space freed up) but that is something I need to do anyway because there is finite capacity in the bandwidth to my radio modem and the radio modem itself has finite capacity which diminishes with distance.  Just another reminder that with embedded systems we can’t have everything we want simultaneously, and the trick is to find a good (and delicate) balance between all the different components and needs.

Update September 7, 2012

Life is never easy!  I have discovered the Overo serial drivers in the Linux 3.x kernel now work a lot harder to match the standard baud rates very closely:  115,200, 230,400, 460,800, and 921,600.  None of these are a close enough match to the APM2′s list of nonstandard baud rates — 200,000, 250,000, 333,333, 500,000, and 1,000,000.  This is a major annoyance!!!  The Overo serial drivers in the Linux 2.6.x kernels seemed to do an actual 500,000 baud, but unfortunately that is gone.  It all has to do with clock dividers and stuff.

What I have done for the moment is work on trimming some of the fat off my binary packets and I’m sending data at 115,200 baud which does (thankfully) still work!  This is sufficient for my current needs so I can move forward, but it does quell any thoughts of running everything end-to-end at 200hz on this layout.  But perhaps with some attention to trimming more fat and perhaps only sending core IMU data at 200hz we could still build a system that inherently has 200hz throughput where it matters.

Introduction

I would like to start off my first post in this series by briefly describing where I’ve been and where I’m going.  Hopefully this will give a bit of context to help understand subsequent posts and hopefully will help put my engineering decisions in some context.

History of the World, Part 1

Somewhere around 2005 I became very interested in aerial robotics.  I have always been an aviation enthusiast, I’ve built model airplanes since I was a kid, and I have been flying RC airplanes since high school.  At the time I was working for the Univ. of Minnesota and managed to get myself sent off to a training class for a little device called a Crossbow MNAV.  The MNAV was a small package that included mems gyros, accelerometers, altimeter, air speed, GPS, RC receiver PPM train decoding, and could control up to 8 servos.  It was perfect for building a small UAV autopilot.  As it turned out, even though the MNAV was pretty good for it’s time, it wasn’t perfect — and about 2 years later when I finally had my code up and flying successfully, Xbow announced they had discontinued the MNAV.

My autopilot code was written for Linux.  I had been running it on a 400mhz Gumstix processor (the really old one) and that talked to the MNAV over a serial connection.  The MNAV sent up the IMU and GPS data to the Gumstix which ran the attitude filter, autopilot, and navigation code, and sent servo positions back to the MNAV.  This worked rather well so I was pretty disappointed when the MNAV was terminated.

History of the World, Part 2

In a world with no MNAV I launched off on an adventure to build my own autopilot from off the shelf parts.  What made this possible was the original ArduPilot.  I could buy IMUs, GPSs, and processors off the shelf, but the ArduPilot gave me the ability to interact with RC servos, the RC receiver, and gave me a hardware failsafe mux, so whenever something went wrong with my code, I could immediately take over manual control of the aircraft.  As it turns out, this has saved me more times than I can count.

I ended up with a system built around the new Gumstix Verdex “pro”.  This processor ran at a whopping 600mhz and had enough number crunching power to run a full blown 15-state kalman filter at 50hz.  I connected a SparkFun 6DOFv4 IMU initially (ultimately upgraded to a VectorNav VN-100) for my inertial sensor, a ublox5 for my GPS, and of course the original ArduPilot for servo control.

At that time I hacked my own version of the ArduPilot firmware to talk directly to the Gumstix Verdex.  I left out all the AP stuff and simply used it to read the RC receiver and relay the pilot stick commands to the Gumstix, and then the Gumstix would send back servo positions to the ArduPilot which would take care of moving the servos to the specified positions.

History of the World Part 3

Time passes, a few projects came and went.  Now it’s 2012, and now the APM2 is released.  Compared to the original ArduPilot, the APM2 is a beautiful work of art!  It looks good, it packs on all kinds of high quality sensors, it has a powerful at mega 2560 processor, it is small, it is light, it is cheap, and it’s pretty well debugged because so many people are buying and flying them.

I decided it would be interesting to revisit my old autopilot design and modernize it based on the new APM2.  Any set of engineering decisions has plusses and minues.  Embedded autopilots are always a delicate balance between, size, weight, cost, capability, power requirements, etc.  And within those larger parameters, every architectural choice has subtle implications related to performance, latency, precision, capability, development tools, integration issues, unexpected surprises (good and bad), and on and on. On the one hand, there is a huge matrix of possibilities, and on the other hand specific choices have cascading effects and lead you in specific directions.  In the end, the goal is to find that magic balance of choices that lead you to an ideal outcome without sacrificing too many things along the way.

Looking to the Future

Where am I going with all this?  What do I envision my autopilot will look like at the end of the process?  Here is a quick summary:

• An APM2 with firmware rewritten to simply collect all the sensor data and pass it up stream to my Gumstix for actual processing.
• A Gumstix Verdex (1Ghz Arm) with hardware floating point, 512Mb of RAM
• Linux based Autopilot software that includes a full 15-state kalman filter, XML configurable PID system, and advanced mission/task system.
• Everything running at 200hz.

I plan to continue this “Hacking the APM2″ series with future posts describing interesting things that I’ve learned or issues that I’ve solved along the path; culminating (hopefully) with a “Super ArduPilot Mega 2 on 1Ghz ARM Steroids” based autopilot system.