A couple of weeks ago we found the NeoPixel ring from Adafruit at a local shop, we had to attach this neat board to our copter and see what cool effect we could make with it. The ring has 16 RGB LEDs that can be driven independently with only one data wire. The LEDs the Neopixel contains are called WS2812.

This post is describing the development of a WS2812 driver for the Crazyflie. A later post will show the usage we did of it (fairly limited in comparison of the seemingly endless possibility, as usual we have more ideas than time to execute them :-).

The WS2812 LED

First of all we needed to see how to control the LEDs. The protocol used by the WS2812 is quite simple but special. All LED on the ring have a data input and a data output pin in a chained manner. Colors of the LED is send over the Data line to the first LED that will save it internally. When sending a second color the first LED will send the saved color to the second LED. And so on, by sending 16 color data we can set the color of all LEDs of the ring independently. Colors are sent as 24 bits (8 bits per component).

Up to there nothing is really peculiar, the system is pretty neat and allows to control a lot of LEDs with just one IO. The problems comes with the Bit encoding:

Bits are encoded with pulse width: a short pulse means 0 and a long pulse means 1. The bitrate is of about 800KHz and the tolerances on bit timing is pretty tight. As there is no hardware peripheral dedicated for this (on common microcontrollers), this bitstream needs to be implemented either by bit-banging or by being a bit creative with the peripheral we actually have at our disposal.

Driver implementation

Adafruit did implement a driver for the WS2812 for Arduino, this is a software implementation that uses the CPU to implement the signal. However at that bitrate a software/bit-bang implementation have to be timed carefully and the easiest for that is actually to implement it in assembler. Arduino runs on an AVR processor and these processor are very predictable: each instruction runs in a specified number of clock cycle which makes it possible to implement a timed loop like the one of Adafruit. However this becomes impractical on bigger CPU that implements caches and other optimization that makes it really hard to predict how much cycle each instruction will take.

The Crazyflie runs a STM32F103 based on an ARM Cortex-M3. This is just complex enough to make the ASM-timed loop impractical: The CPU core runs at 72MHz but the flash is slower so a simple cache memory is inserted in the middle and depending of the state of this cache it may take from 1 to 3 cycles to execute an instruction. There is an even bigger problem: a CPU timed loop requires to stop all interrupt and to have the CPU running exclusively on updating the WS2812. We cannot allow that on the Crazyflie that require a 250Hz control loop to stay (controllably) airborne.

We asked Google to see if someone already came with a solution and actually someone did. The Elia’s Electronics Blog posted a neat, well documented, solution using the STM32 Pulse Width Modulation (PWM) capabilities of the STM32 to generate signals that the WS2812 understands. All we needed was a timer/pwm output then. It happens that we have one timer1 output on the Crazyflie extension port:

This is only 3 wires: VCOM for the battery voltage, GND and the timer output.

I started porting the Elia’s code to the Crazyfle. The only free timer we could easily use was different and much more complex than the one Elia uses. So part of the frustrating implementation was to figure out WHY the signal was looking weird on the scope! Finally after an hour or so the code was working:

Enhancements

Now that was not quite enough. The driver is using the PWM to generate the 1’s and 0’s pulse width and the DMA is used to feed the bit width independently of the CPU (so that it can do something else, like controlling the copter attitude…). It requires all the 16 LEDs data to be written in the memory buffer before starting the DMA. Each LED has 24 bits color data and each bit will be encoded as 16bit pulse length for the timer. It means that the full ring will take 768Bytes in RAM. It sounds small but its a bit too much for us and, most importantly, it does not scale: if we want a second ring the ram requirement will double.

To fix this I implemented two things: An easy one is that the DMA can do some type conversion and one of these is that it can read 8 bit in memory and write 16 bit in the peripheral. This allows to store only 8 bits pulse width in memory and so divides by 2 the memory requirement, but it still doesn’t scale.

The fix to scaling is double buffering. The STM32F103 does not implement DMA double buffering as such but what it has is close enough: circular DMA with half-transfer interrupt. The idea is to load the 2 first LED in a buffer and to start the DMA for 42 bytes in circular mode. When the DMA has transferred the first LED it triggers the half-transfers interrupt which allows the CPU to replace the first LED data by the 3rd. When the 2nd LED is transferred the transfer-complete interrupt is triggered by the DMA and the CPU fills in the 4th LED in place of the 2nd one. The DMA roll-over at the beginning of the buffer and sends what is now the 3rd LED data. This continues until all the LEDs has been  sent. This solution scales: it requires a 42 bytes buffer for as many LED as we want!

Conclusion

All that process was not really simple. However the result is simple, now the only thing required to control a Neopixel ring in the Crazyflie is:

#define BLACK {0x00, 0x00, 0x00}</span>
static uint8_t color[][3] = {{40, 40, 40}, {32, 32, 32}, {16,16,16}, {8,8,8},
                             {4,4,4}, {2,2,2}, {1,1,1}, BLACK,
                             BLACK, BLACK, BLACK, BLACK,
                             BLACK, BLACK, BLACK, BLACK,
                            };

/* ... */
ws2812Send(color, 16);

And it will run nicely without interrupting other tasks like control and communication. Replace 16 by 32 to control 2 Neopixel rings. The current implementation has been pushed in the crazyflie-firmware neopixel_dev branch (still require some clean-ups!). Actual implementation is in ws2812.c and neopixelring.c.

The only problem left is not software: At the end of the battery life, the voltage is not enough to fully lit the blue LED. That make all white look orange-ish. The only way to fix this problem would be to step-up the battery voltage higher then the forward voltage of the blue LED. Though it works well enough without that.

Stay tuned for a following post about actual usage of the ring.

A couple of weeks ago we attached a uBlox MAX-7 GPS module to the Crazyflie (blog post). Back then it was mostly a proof of concept, all we did was to re-route the raw GPS data (in text NMEA format) directly to the PC using the Crazyflie text console port. This allowed us to quickly prove that a GPS can work on the Crazyflie but was not that useful and efficient: the copter did not decode the gps position and a lot of radio bandwidth was used. Last Friday we decided to fix it and to make it clean(er).

The ultimate goal was to measure the Crazyflie speed, if it wasn’t for the rain we could have done the measurement! Anyway, this work allowed us to exercise the debug functionality of the Crazyflie platform and so to see the strength of it but also what needs to be enhanced. In this post we will try to document (at high level) the steps taken to implement the GPS in the Crazyflie. The source code is pushed in the crazyflie firmware and python-client git repos. The Python client code is in the master branch and the firmware code in the gpu_ublox_dev branch (dev branch means that the code is far from final/clean, but it works!).

Electronically the GPS is connected using only 4 pins: VCC, GND, serial RX and serial TX. The serial port is connected to pins 3 and 5 of the expansion header. The power is connected to VCC.

The electronic was already tested and working so we had 2 tasks left:

• Decoding the GPS information in the firmware and creating log variables to make the data available for the PC software
• Updating the GPS tab of the PC software to fetch GPS data from the log subsystem instead of parsing it from the text console

It happens that these two tasks could be done mostly independently and Marcus and I started to work in parallel. The only thing we had to agree upon was which log variable and what scaling to use for the variables. We used the format that the GPS chip is already using which made things easier.

Firmware

For the firmware part, the first step was to acquire the GPS data. GPS chips usually can talk two languages: the standard text-based NMEA and some kind of proprietary binary format, UBX for uBlox. I chose the binary format as it is a lot easier use in C: no text parsing has to be done, all data dirrectly fits in a C structure. But first the GPS has to be setup to output data in binary modes and to output the data we were interested in. To quickly setup the GPS I used a tool that uBlox provides and that permits to generate proper UBX messages:

Two UBX messages where required: One to disable NMEA output and one to enable the NAV-PVP message which contains basically all data you would want from a GPS (position, speed, date and the accuracy). Once this is sent to the GPS chip it starts to send a NAV-PVP UBX packet once a second. Then, the GPS acquisition loop in the Crazyflie (currently implemented in the UART task, so it has to be moved into a proper driver) just has to wait for an UBX packet, read it and if it is a NAV-PVT packet then extract values from it. The GPX code has been tested on PC using another uBlox receiver connected to a USB serial cable. Then after copying the newly added uartReceiveUbx() function  into the Crazyflie firmware, the GPS acquisition loop looks like this:

  while(1)
  {
    uartReceiveUbx(&msg, 100);

    if (msg.class_id == NAV_PVT) {
      gps_fixType = msg.nav_pvt->fixType;
      gps_lat = msg.nav_pvt->lat;
      gps_lon = msg.nav_pvt->lon;
      gps_hMSL = msg.nav_pvt->hMSL;
      gps_hAcc = msg.nav_pvt->hAcc;
      gps_gSpeed = msg.nav_pvt->gSpeed;
      gps_heading = msg.nav_pvt->heading;
    }

    ledseqRun(LED_GREEN, seq_linkup);
  }

The last thing, to make the data available from the PC, is to add a GPS log block and to add variables to it:

LOG_GROUP_START(gps)
LOG_ADD(LOG_UINT8, fixType, &gps_fixType)
LOG_ADD(LOG_INT32, lat, &gps_lat)
LOG_ADD(LOG_INT32, lon, &gps_lon)
LOG_ADD(LOG_INT32, hMSL, &gps_hMSL)
LOG_ADD(LOG_UINT32, hAcc, &gps_hAcc)
LOG_ADD(LOG_INT32, gSpeed, &gps_gSpeed)
LOG_ADD(LOG_INT32, heading, &gps_heading)
LOG_GROUP_STOP(gps)

Debugging the firmware code can be done using the client log plotter tab, not so nice to look at positioning but good enough to see if it is working (when the accuracy, gps.hAcc,  goes down the GPS has a fix!):

Client

Once the log block was decided, Marcus could start the client development by updating the debug driver. The debug link is a module of the Python client that behaves like a CRTP (the Crazyflie protocol) link but is in fact just some software running offline. It allows to easily develop and debug the client without requiring the usage of a Crazyflie. The debug link was modified to include all variable that the Crazyflie would eventually contain and to give them some value that can be logged.

When this was done, the GPS tab has to be updated to display actual values, a max speed and reset button is also added (the idea was to measure the Crazyflie speed). After fighting more than expected with the QT layouts the result is good enough:

Note that the client uses the Python binding of KDE Marble which has to be compiled manually. Only Marcus has had the courage to do that on his computer, but luckily he also compiled it on the latest Bitcraze VM so that we can all easily enjoy the new GPS tab :).

Merge

Now that the client and the firmware are made separately we ‘just’ have to connect the new client to the new firmware. And guess what? it worked the first time :-) (Yes I know you have no reason to believe me but this time it really worked the first time).

Unfortunately for us last Friday was one of these Swedish rainy day, all we could do was to take turns to stand in the middle of the road outside of our office, in the rain, holding the Crazyflie in a plastic bag and waiting for the GPS to get a fix (people passing by were looking quite strangely at us …). It happens that the rain where not helping at all! And the fact that we don’t have assisted-GPS (yet) means that the GPS would get a fix in 40sec best case, it took about 5-10minutes for us. But eventually we got the fix:

Conclusion

One thing we have to work on is the modularity of the firmware. Things like having a clear and easy to use HAL for peripherals on the extension port. It is on our ToDo list and it would have been useful here to do a cleaner job with the firmware implementation. A good thing is that while this implementation is uBlox specific for the firmware part, it is completely hardware-independent on the client side. It means that it is possible to implement any kind of positioning, with other GPS chip or other technology, and as long as this positioning declares the right log variable the client will work with it unmodified.

As for the GPS, uploading assistance data to the Crazyflie would permit to drastically reduce the fix time to about 10-15sec. Also this GPS is capable of 10Hz update rate which would be nice to test. The GPS on a Crazyflie is still mostly a proof-of-concept and is of course not useful for indoor flight. Though with light winds the Crazyflie is pretty capable outdoor, so with GPS capability it could be interesting to experiment a bit with trajectory planing. Of course this is even more true for country with a warm and dry weather :-).