openpilot is software that produces serial messages (CAN Messages) to change the acceleration and steering angle of a car given some camera streams, existing serial messages from the car, and sensor data. In addition to the real-time messages, openpilot produces logs that are used to train machine learning models at a later date.

The goal of this introduction is to introduce you to the moving pieces of openpilot and help you understand how they work together to create these outputs.

Modern vehicles communicate using CAN messaging, a broadcast protocol that allows many computers to talk together in a way that is tolerant to noisy environments. From the openpilot perspective, the good thing about the CAN protocol is that it is inherently trusting, allowing messages to be spoofed. The bad thing about the CAN protocol is that each manufacturer creates their own dictionary of CAN message IDs and data definitions.

For example, one manufacturer might put the steering angle on message ID 0x30 at bytes 3 and 4, while another manufacturer describes the steering angle on message ID 0xe4 at bytes 2 and 3.

So even after openpilot figures out what the acceleration and steering angle should be, there is still much work to be done to turn it into a CAN message that will be understood by that particular make/model vehicle.

Below is the heart of the code for turning calculations, CC, into manufacturer-specific CAN messages using the car interface, CI, which contains information about the make/model of the car:

# selfdrive/controls/controlsd.py

# send car controls over can
can_sends = CI.apply(CC)
pm.send('sendcan', can_list_to_can_capnp(can_sends))

On each loop of the controlsd process, the car control message, CC, represents the desired setpoints of the car (acceleration, steering angle, etc). The CI.apply method turns these setpoints into make/model specific CAN messages.

On the next line, the pm.send function publishes the can_sends messages on the sendcan topic, in Cap'n Proto format. The next snippet of code describes what happens to the message next. Don't worry if it doesn't make sense yet.

// selfdrive/boardd/boardd.cc

// subscribe to the sendcan topic
SubSocket * subscriber = SubSocket::create(context, "sendcan");

...

while (panda->connected) {
    // get messages as fast as possible
    Message * msg = subcriber->receive();

    // parse the cap'n proto messages
    cereal::Event::Reader event = cmsg.getRoot<cereal::Event>();

    // send the unpacked CAN messages using the panda interface
    panda->can_send(event.getSendcan());
}

The boardd process subscribes to the sendcan topic and turns Cap'n Proto messages into physical CAN messages through the panda hardware and firmware. The panda->can_send method writes to the hardware's microcontroller, which uses CAN transceivers to send CAN messages directly to the vehicle's CAN network.

To recap: controlsd is a process that takes many inputs and turns them into a list of CAN messages for a particular make/model vehicle, and boardd is a process that talks to the vehicle through the panda interface to send (and receive) physical CAN messages. The two processes communicate with each other using a pub-sub messaging framework called cereal where pm denotes that a process is a topic publisher, and sm denotes that a process is a topic subscriber.

This separation of concerns into purposeful processes that communicate through pub-sub messaging will be covered in greater detail in the next section.

So far, we've mentioned two processes: controlsd and boardd. This section will give a brief overview of the major processes and how they communicate with each other. You may have noticed that processes all end with the letter d. That's a nod to Linux daemons, background processes that have no user control. Similarly, openpilot's processes run without any direct user intervention.

• camerad: interfaces with available camera hardware to generate images, which are encoded as videos
• sensorsd: interfaces with available sensor hardware to generate data from the hardware's IMU
• boardd: interfaces with vehicle's CAN network and panda's ublox GPS chip to send and generate data
• ubloxd: converts raw GPS messages into real-time GPS data
• locationd: combines inputs from sensors, GPS, and model into real-time estimates for movement and position
• modeld: transforms camera images into estimates for movement, future position, and the location of other vehicles
• plannerd: TODO
• controlsd: combines a wide range of inputs to produce car-specific messages for a desired future state
• paramsd: TODO
• thermald: monitors the vitals of the hardware running openpilot
• loggerd: logs and uploads messages history and videos
• athenad: opens a websocket connection so that message history and videos can be uploaded to the cloud

In the openpilot process map below, processes are represented by nodes, and the topics that they publish-to and subscribe-to are represented by the arrows between the nodes. A process is a publisher of a topic if the arrow points away from the process and is a subscriber if the arrow points toward the process. So, for example, the thermald process publishes deviceState and subscribes to managerState, pandaState, and gpsLocationExternal.

In the cereal pub-sub framework, any process can subscribe to any topic, but each topic can only have one publisher. Messages are stored and exchanged in Cap'n Proto format, an extremely fast and lightweight way to send objects (like JSON) as binary (like ProtoBufs). The structure of every message is contained in the log.capnp file in the cereal sub-module.

In the log.capnp file, a single Event struct contains all the topic message definitions:

# cereal/log.capnp

struct Event {
  logMonoTime @0 :UInt64;  # nanoseconds
  ...
  union {
    ...
    # ********** openpilot daemon msgs **********
    can @5 :List(CanData); # CAN received
    controlsState @7 :ControlsState;
    sensorEvents @11 :List(SensorEventData);
    pandaState @12 :PandaState;
    radarState @13 :RadarState;
    liveTracks @16 :List(LiveTracks);
    sendcan @17 :List(CanData); # CAN sent
    liveCalibration @19 :LiveCalibrationData;
    carState @22 :Car.CarState;
    ...
  }
}

Notice how both can and sendcan use the type List(CanData). List is a primitive type to Cap'n Proto, and CanData is a struct that's also defined in log.capnp. The CanData struct is composed exclusively of primitive types, so no further definition is needed:

# cereal/log.capnp

struct CanData {
  address @0 :UInt32; # 11 or 29-bit address
  busTime @1 :UInt16;
  dat     @2 :Data; # up to 8 bytes of data
  src     @3 :UInt8; # 0-3 or 128-131
}

In C++, accessing a property foo from Cap'n Proto structs is as easy as calling getFoo(). If foo is a struct, getFoo() returns a Foo::Reader, a read-only class. Otherwise, if foo is a primitive, getFoo() returns the primitive value. To get the top-level Event struct, we called a special method getRoot(), which is cast as the root type cereal::Event:

// selfdrive/boardd/boardd.cc

// get the root struct
cereal::Event::Reader event = cmsg.getRoot<cereal::Event>();
// pass the sendcan struct to the panda->can_send method
panda->can_send(event.getSendcan());

Calling event.getSendcan() therefore returns a List<cereal::CanData>::Reader, because List is primative but CanData a custom struct. Then calling getAddress() on the Reader, returns a value.

// selfdrive/boardd/panda.cc

void Panda::can_send(capnp::List<cereal::CanData>::Reader can_data_list) {

    const int msg_count = can_data_list.size();
    for (int i = 0; i < msg_count; i++) {
        // access structs in a list
        auto cmsg = can_data_list[i];
        // get the properties of the struct
        uint32_t address = cmsg.getAddress();
        uint8_t src = cmsg.getSrc();
    }
    ...
    usb_bulk_write(3, (unsigned char*)send.data(), send.size(), 5);
}

In python, reading Cap'n Proto messages is much less verbose. Here we read the deviceState messages:

# selfdrive/controls/controlsd.py

import cereal.messaging as messaging
...
sm = messaging.SubMaster(['deviceState', ...)
...

while True:
    # Create events for battery, temperature, disk space, and memory
    if EON and sm['deviceState'].batteryPercent < 1 and sm['deviceState'].chargingError:
      # at zero percent battery, while discharging, OP should not allowed
      events.add(EventName.lowBattery)
    if sm['deviceState'].thermalStatus >= ThermalStatus.red:
      events.add(EventName.overheat)
    if sm['deviceState'].freeSpacePercent < 7 and not SIMULATION:
      # under 7% of space free no enable allowed
      events.add(EventName.outOfSpace)

Publishing Cap'n'proto messages is just as easy. Here we see how paramsd uses cereal's messaging library to publish a message on the liveParameters topic every time it receives a message on the liveLocationKalman topic:

# selfdrive/locationd/paramsd.py
import cereal.messaging as messaging

# subscribe to a list of topics
sm = messaging.SubMaster(['liveLocationKalman', 'carState'], poll=['liveLocationKalman'])
# publish a list of topics
pm = messaging.PubMaster(['liveParameters'])

while True:
    # when the liveLocationKalman receives a new message
    if sm.updated['liveLocationKalman']:

      msg = messaging.new_message('liveParameters')
      msg.logMonoTime = sm.logMonoTime['carState']
      ...

      msg.liveParameters.posenetValid = True
      msg.liveParameters.sensorValid = True
      msg.liveParameters.steerRatio = float(x[States.STEER_RATIO])
      msg.liveParameters.stiffnessFactor = float(x[States.STIFFNESS])
      msg.liveParameters.angleOffsetAverageDeg = angle_offset_average
      msg.liveParameters.angleOffsetDeg = angle_offset
       msg.liveParameters.valid = all((
        abs(msg.liveParameters.angleOffsetAverageDeg) < 10.0,
        abs(msg.liveParameters.angleOffsetDeg) < 10.0,
        0.2 <= msg.liveParameters.stiffnessFactor <= 5.0,
        min_sr <= msg.liveParameters.steerRatio <= max_sr,
      ))
      ....
      pm.send('liveParameters', msg)

Notice how the Cap'n Proto struct from cereal matches the message properties, but not all properties are required, such as gyroBias. In the case that properties are not defined, they will be set to 0 or false:

# cereal/log.capnp

liveParameters @61 :LiveParametersData;

...

struct LiveParametersData {
  valid @0 :Bool;
  gyroBias @1 :Float32; #
  angleOffsetDeg @2 :Float32;
  angleOffsetAverageDeg @3 :Float32;
  stiffnessFactor @4 :Float32;
  steerRatio @5 :Float32;
  sensorValid @6 :Bool;
  yawRate @7 :Float32;
  posenetSpeed @8 :Float32;
  posenetValid @9 :Bool;
}

Now that we understand what the major processes are and how they talk to each other, let's take a look at how data is persisted for time travel debugging and machine learning.

There are two types of log files in openpilot. Cap'n Proto logs, and videos. The two are used in conjunction to replay drives, diagnose problems, train machine learning models, and reverse engineer CAN message definitions. In this section, we'll focus on the Cap'n Proto logs and leave video for later in our introduction. So what is a Cap'n Proto log file?

To answer that question, we'll return our Event struct from cereal's log.capnp file.

struct Event {
  logMonoTime @0 :UInt64;  # nanoseconds
  valid @63 :Bool = true
  ...

  union {
    ...
    gpsNMEA @3 :GPSNMEAData;
    can @5 :List(CanData);
    controlsState @7 :ControlsState;
    sensorEvents @11 :List(SensorEventData);
    ...
  }
}

The important thing to understand here is that union is like an enum, meaning that each instance of Event contains exactly one of the message definitions defined in the union. Thus, each Event has a logMonoTime, a valid property, and a message.

Now that we know what the Event struct is, let's have a look at 20ms of decompressed log data converted to JSON. You may notice that the log is just a series of Event instances:

[{
   "LogMonoTime":"31132127675929",
   "Valid":true,
   "SensorEvents":[
      {
         "Version":104,
         "Sensor":5,
         "Type":16,
         "Timestamp":"31132121969660",
         "UncalibratedDEPRECATED":false,
         "GyroUncalibrated":{
            "V":[
               -0.0488739013671875,
               0,
               0.0122222900390625,
               -0.05096435546875,
               0.00665283203125,
               0.013031005859375
            ],
            "Status":0
         },
         "Source":0
      },
      {
         "Version":104,
         "Sensor":1,
         "Type":1,
         "Timestamp":"31132121969660",
         "UncalibratedDEPRECATED":false,
         "Acceleration":{
            "V":[
               9.889938354492188,
               0.772552490234375,
               0.090179443359375
            ],
            "Status":3
         },
         "Source":0
      }
   ]
},
{
   "LogMonoTime":"31132130390148",
   "Valid":true,
   "Can":[
      {
         "Address":404,
         "BusTime":23453,
         "Dat":"EQAXAAKEWKM=",
         "Src":1
      },
      {
         "Address":405,
         "BusTime":23573,
         "Dat":"EQAAAAaEWJE=",
         "Src":1
      },
      ...
   ]
},
{
   "LogMonoTime":"31132135487491",
   "Valid":true,
   "RoadCameraState":{
      "FrameId":1,
      "EncodeId":0,
      "TimestampEof":"31132109621000",
      "FrameLength":5419,
      "IntegLines":5408,
      "GlobalGain":510,
      "LensPos":635,
      "LensSag":0,
      "LensErr":0,
      "LensTruePos":400,
      ...
   }
},

This is only a very small snippet. During normal operation thousands of messages are sent per second. Every message sent between processes is saved to a raw log file, called an rlog. This file is uploaded to comma server's (with the user's permission) when the device is able to connect to WiFi.

With so much data being saved and sent, the name of the game is compression. There are two main ways that openpilot solves the data compression problem. The first is through the data format, and the second is through sampling rates.

When it comes to the data format, openpilot uses Cap'n Proto for its small in-memory footprint, language interoperability, and lightning speed. For compression, we use the bzip algorithm, which favors smaller size over speed.

Let's now take a look at how the loggerd process saves all messages to disk in bz2 format:

// selfdrive/loggerd/loggerd.cc

Poller * poller = Poller::create();

  // iterate through all the services (topics)
  for (const auto& it : services) {
    // subscribe to every topic
    SubSocket * sock = SubSocket::create(s.ctx, it.name);
    poller->registerSocket(sock);
    ...
  }

  // poll for new messages on all topics
  for (auto sock : poller->poll(1000)) {
      while (msg = sock->receive(true)) {
          // decide whether the message should be logged in the qlog
          const bool in_qlog = qs.freq != -1 && (qs.counter++ % qs.freq == 0);
          // save the message to disk
          logger_log(&s.logger, (uint8_t *)msg->getData(), msg->getSize(), in_qlog);
      }
  }

The loggerd process subscribes to every message, and calls logger_log on the message data, which eventually uses libbz2 to compress the byte array message into bz2 format, and write it to disk.

// selfdrive/loggerd/logger.h

inline void write(void* data, size_t size) {
    ...
    do {
        ...
        BZ2_bzWrite(&bzerror, bz_file, data, size);
    } while (bzerror == BZ_IO_ERROR && errno == EINTR);
    ...
}

We've just described how an rlog.bz2 file is created. But the careful reader may have noticed something called a qlog. When loggerd calls logger_log it passes an argument called in_qlog. The qlog.bz2 file is a subset of rlog.bz2. Instead of saving all messages, the qlog file samples every n-th message as defined by the cereal services file:

# cereal/services.py

services = {
  # service: (should_log, frequency, qlog decimation (optional))
  "sensorEvents": (True, 100., 100),
  "gpsNMEA": (True, 9.),
  "deviceState": (True, 2., 1),
  "can": (True, 100.),
  "controlsState": (True, 100., 10),
  "pandaState": (True, 2., 1),
  "radarState": (True, 20., 5),
  "roadEncodeIdx": (True, 20., 1),
  "liveTracks": (True, 20.),
  "sendcan": (True, 100.),
  "logMessage": (True, 0.),
  "liveCalibration": (True, 4., 4),
  ...
}

The services dictionary specifies whether the messages from that service should be logged, how many times the message is sent per second (Hz), and the sampling rate for qlog, called decimation.

Let's look at a few examples to see how this works. The deviceState message is sent 2 times per second (2Hz), and every message is saved to the qlog. The radarState message is sent 20 times per second, and one out of every 5 messages is saved to the qlog. The can message is sent 100 times per second, but no messages are saved to the qlog. The purpose of the qlog is for streaming real-time data for users with a cellular data connection. Because cellular data is more expensive, the data resource is limited.

Before moving on to the next section, let's revisit the code for deciding whether a message should be in qlog:

// selfdrive/loggerd/loggerd.cc

  // iterate through the services
  for (const auto& it : services) {
    // subscribe to all messages topics
    ...
    // store the message sampling rates for qlog
    qlog_states[sock] = {.counter = 0, .freq = it.decimation};
  }

  // poll for new messages on all topics
  for (auto sock : poller->poll(1000)) {

      QlogState &qs = qlog_states[sock];

      while (msg = sock->receive(true)) {
          // decide whether the message should be logged in the qlog
          const bool in_qlog = qs.freq != -1 && (qs.counter++ % qs.freq == 0);
          // save the message to disk
          logger_log(&s.logger, (uint8_t *)msg->getData(), msg->getSize(), in_qlog);

          // check to see if we need a new log file
          rotate_if_needed();
      }
  }

Now, that we understand how rlog.bz2 and qlog.bz2 files are created, we need to give some thought to how to break up the files into manageable chunks. To do this, we use the function rotate_if_needed seen in the code above.

The purpose of the rotate_if_needed function is to break log files up into 60-second chunks, called segments. Because all log files are named rlog.bz2 or qlog.bz2, we need some means of differentiating between them. To accomplish this, we use folders with the following naming convention <LOG_ROOT>/<ROUTE>--<SEGMENT>/.

LOG_ROOT is defined as:

// selfdrive/hardware.h

inline std::string log_root() {
    return Hardware::PC() ? HOME + "/.comma/media/0/realdata" : "/data/media/0/realdata";
}

ROUTE is a date/timestamp that is defined once on power-up. It looks like 2021-09-16--19-49-40, which is generated by the following function:

// selfdrive/loggerd/logger.cc

std::string logger_get_route_name() {
  char route_name[64] = {'\0'};
  time_t rawtime = time(NULL);
  struct tm timeinfo;
  localtime_r(&rawtime, &timeinfo);
  strftime(route_name, sizeof(route_name), "%Y-%m-%d--%H-%M-%S", &timeinfo);
  return route_name;
}

Finally, SEGMENT is a counter variable, converted to a string. The function rotate_if_needed eventually increments the segment number when a new segment is needed. By default, segments are 60 seconds long, such that segment 42 represents the 42nd minute of the drive. Thus, a valid path for an rlog of the 42nd minute of a drive on 9/16/2021 could be: /data/media/0/realdata/2021-09-16--19-49-40--42/rlog.bz2.

While messages provide access to real-time state, and logs give us the ability to look back in time, sometimes we need to persist global data between processes. For example, we don't want to ask the user whether to upload logs every time we restart the device or connect to WiFi. It's better to save the user selection as a persistent parameter.

For a lot of developers, persistent key/value storage might sound like a job for a database. But remote databases are not an option because the hardware cannot always access the internet, and local databases like sqlite contain unnecessary complexity for our simple needs. In order to run many processes at near real-time speeds, everything must be optimized for speed and efficiency.

Thus, the method employed by openpilot for persisting key/value data is to use the filesystem, using a library called params. Params stores keys/values on the Linux filesystem here:

// selfdrive/hardware.h

inline std::string params() {
  return Hardware::PC() ? HOME + "/.comma/params" : "/data/params";
}

We haven't talked about Hardware::PC() yet, but you can probably guess that it is a method to identify whether we're using a PC (without Android) or not. Great! Now let's take a look at some of the keys that we'll persist and the policies for when to clear the values:

// selfdrive/common/params.cc

std::unordered_map<std::string, uint32_t> keys = {
    {"AccessToken", CLEAR_ON_MANAGER_START | DONT_LOG},
    {"ApiCache_DriveStats", PERSISTENT},
    {"ApiCache_Device", PERSISTENT},
    {"ApiCache_Owner", PERSISTENT},
    {"ApiCache_NavDestinations", PERSISTENT},
    {"AthenadPid", PERSISTENT},
    {"CalibrationParams", PERSISTENT},
    {"CarBatteryCapacity", PERSISTENT},
    {"CarParams", CLEAR_ON_MANAGER_START | CLEAR_ON_PANDA_DISCONNECT | CLEAR_ON_IGNITION_ON},
    {"CarParamsCache", CLEAR_ON_MANAGER_START | CLEAR_ON_PANDA_DISCONNECT},
    {"CarVin", CLEAR_ON_MANAGER_START | CLEAR_ON_PANDA_DISCONNECT | CLEAR_ON_IGNITION_ON},
    {"CommunityFeaturesToggle", PERSISTENT},
    {"ControlsReady", CLEAR_ON_MANAGER_START | CLEAR_ON_PANDA_DISCONNECT | CLEAR_ON_IGNITION_ON},
    {"CurrentRoute", CLEAR_ON_MANAGER_START | CLEAR_ON_IGNITION_ON},
    {"DisableRadar", PERSISTENT}, // WARNING: THIS DISABLES AEB
    {"EndToEndToggle", PERSISTENT},
    {"CompletedTrainingVersion", PERSISTENT},
    {"DisablePowerDown", PERSISTENT},
    {"DisableUpdates", PERSISTENT},
    {"EnableWideCamera", CLEAR_ON_MANAGER_START},
    {"DoUninstall", CLEAR_ON_MANAGER_START},
    {"DongleId", PERSISTENT},
    ...
    {"GsmRoaming", PERSISTENT},
    {"HardwareSerial", PERSISTENT},
    {"HasAcceptedTerms", PERSISTENT},
    {"IMEI", PERSISTENT},
    {"InstallDate", PERSISTENT},
    {"IsDriverViewEnabled", CLEAR_ON_MANAGER_START},
    {"IsLdwEnabled", PERSISTENT},
    {"IsMetric", PERSISTENT},
    {"IsOffroad", CLEAR_ON_MANAGER_START},
    {"IsOnroad", PERSISTENT},
    {"IsRHD", PERSISTENT},
    {"IsTakingSnapshot", CLEAR_ON_MANAGER_START},
    {"IsUpdateAvailable", CLEAR_ON_MANAGER_START},
    {"UploadRaw", PERSISTENT},
    ...
    {"JoystickDebugMode", CLEAR_ON_MANAGER_START | CLEAR_ON_IGNITION_OFF},
};

The params library has two primary methods: Params().get(key) and Params().put(key, value). The Params().get(key) method is blocking, and Params.put(key, value) writes atomically. Let's take a look at how these methods are used to cache the expensive operation of determining the car firmware and VIN number:

# selfdrive/car/car_helpers.py

# try to load the car params from the cache
cached_params = Params().get("CarParamsCache")
if cached_params is not None:
    # use the cap'n proto struct to convert the cached bytes into a python object
    cached_params = car.CarParams.from_bytes(cached_params)

if cached_params is not None and len(cached_params.carFw) > 0 and cached_params.carVin is not VIN_UNKNOWN:
    # use the cached params
    vin = cached_params.carVin
    car_fw = list(cached_params.carFw)
else:
    # do expensive operation to get vin and firmware with CAN bus messaging
    vin = get_vin(logcan, sendcan, bus)
    car_fw = get_fw_versions(logcan, sendcan, bus)

...
# persist the car's VIN number
Params().put("CarVin", vin)

In the example above, we check our persistent storage for the carParamsCache key. If the value is present, we avoid doing the expensive get_vin and get_fw_versions function calls.

In addition to get and put, the params library has the methods get_bool and put_bool, which are used for true/false values. In C++, the methods are getBool and putBool. In the next example, we see how the ControlsReady param is used to tell the boardd process to wait for the controlsd process to finish loading the car parameters:

# selfdrive/controls/controlsd.py

# after fingerprinting and loading the car interface
Params().put_bool("ControlsReady", True)

// selfdrive/boardd/boardd.cc
p = Params()
...
// wait for the controls ready param to be set
while (true) {
    ...
    if (p.getBool("ControlsReady")) {
      params = p.get("CarParams");
      if (params.size() > 0) break;
    }
    util::sleep_for(100);
  }

This code is called when the boardd process starts. It waits for the CarParams to be set so that it can configure the panda's "safety hooks" according to the make/model of the car. Recall from list of keys that ControlsReady has a policy of CLEAR_ON_MANAGER_START | CLEAR_ON_PANDA_DISCONNECT | CLEAR_ON_IGNITION_ON. That means that any time the device, car, or panda is disconnected, the parameter will be cleared.

In the next section, we'll learn about the fingerprinting process, car interfaces, and manufacturer-specific safety hooks.

So far, we've looked at the mechanics of how data is transported in openpilot. In the next few sections, we'll try to understand the interfaces required to support self-driving in hundreds of different cars. To do that, let's start by defining what a fingerprint is.

In openpilot, a fingerprint is a dictionary of CAN message IDs and data length (in bytes). A fingerprint is used to identify a car based on the set of CAN messages sent over a few seconds while ignoring the content of the messages.

Now suppose there are only two cars, XTRAIL and LEAF, in the universe. And suppose that each car can have two different possible fingerprints, depending on some manufacturing variability (like the model year). Two cars with two fingerprints give us four possible fingerprints:

# selfdrive/car/nissan/values.py
fingerprints = {
  CAR.XTRAIL: [
    {
      2: 5, 42: 6, 346: 6, 347: 5, 348: 8, 349: 7, 361: 8, 386: 8, 389: 8, 397: 8, 398: 8, 403: 8, 520: 2, 523: 6, 548: 8, 645: 8, 658: 8, 665: 8, 666: 8, 674: 2, 682: 8, 683: 8, 689: 8, 723: 8, 758: 3, 768: 2, 783: 3, 851: 8, 855: 8, 1041: 8, 1055: 2, 1104: 4, 1105: 6, 1107: 4, 1108: 8, 1111: 4, 1227: 8, 1228: 8, 1247: 4, 1266: 8, 1273: 7, 1342: 1, 1376: 6, 1401: 8, 1474: 2, 1497: 3, 1821: 8, 1823: 8, 1837: 8, 2015: 8, 2016: 8, 2024: 8
    },
    {
      2: 5, 42: 6, 346: 6, 347: 5, 348: 8, 349: 7, 361: 8, 386: 8, 389: 8, 397: 8, 398: 8, 403: 8, 520: 2, 523: 6, 527: 1, 548: 8, 637: 4, 645: 8, 658: 8, 665: 8, 666: 8, 674: 2, 682: 8, 683: 8, 689: 8, 723: 8, 758: 3, 768: 6, 783: 3, 851: 8, 855: 8, 1041: 8, 1055: 2, 1104: 4, 1105: 6, 1107: 4, 1108: 8, 1111: 4, 1227: 8, 1228: 8, 1247: 4, 1266: 8, 1273: 7, 1342: 1, 1376: 6, 1401: 8, 1474: 8, 1497: 3, 1534: 6, 1792: 8, 1821: 8, 1823: 8, 1837: 8, 1872: 8, 1937: 8, 1953: 8, 1968: 8, 2015: 8, 2016: 8, 2024: 8
    },
  ],
  CAR.LEAF: [
    {
      2: 5, 42: 6, 264: 3, 361: 8, 372: 8, 384: 8, 389: 8, 403: 8, 459: 7, 460: 4, 470: 8, 520: 1, 569: 8, 581: 8, 634: 7, 640: 8, 644: 8, 645: 8, 646: 5, 658: 8, 682: 8, 683: 8, 689: 8, 724: 6, 758: 3, 761: 2, 783: 3, 852: 8, 853: 8, 856: 8, 861: 8, 944: 1, 976: 6, 1008: 7, 1011: 7, 1057: 3, 1227: 8, 1228: 8, 1261: 5, 1342: 1, 1354: 8, 1361: 8, 1459: 8, 1477: 8, 1497: 3, 1549: 8, 1573: 6, 1821: 8, 1837: 8, 1856: 8, 1859: 8, 1861: 8, 1864: 8, 1874: 8, 1888: 8, 1891: 8, 1893: 8, 1906: 8, 1947: 8, 1949: 8, 1979: 8, 1981: 8, 2016: 8, 2017: 8, 2021: 8, 643: 5, 1792: 8, 1872: 8, 1937: 8, 1953: 8, 1968: 8, 1988: 8, 2000: 8, 2001: 8, 2004: 8, 2005: 8, 2015: 8
    },
    # 2020 Leaf SV Plus
    {
      2: 5, 42: 8, 264: 3, 361: 8, 372: 8, 384: 8, 389: 8, 403: 8, 459: 7, 460: 4, 470: 8, 520: 1, 569: 8, 581: 8, 634: 7, 640: 8, 643: 5, 644: 8, 645: 8, 646: 5, 658: 8, 682: 8, 683: 8, 689: 8, 724: 6, 758: 3, 761: 2, 772: 8, 773: 6, 774: 7, 775: 8, 776: 6, 777: 7, 778: 6, 783: 3, 852: 8, 853: 8, 856: 8, 861: 8, 943: 8, 944: 1, 976: 6, 1008: 7, 1009: 8, 1010: 8, 1011: 7, 1012: 8, 1013: 8, 1019: 8, 1020: 8, 1021: 8, 1022: 8, 1057: 3, 1227: 8, 1228: 8, 1261: 5, 1342: 1, 1354: 8, 1361: 8, 1402: 8, 1459: 8, 1477: 8, 1497: 3, 1549: 8, 1573: 6, 1821: 8, 1837: 8
    },
  ],
}

The fingerprint values are a set of key/value pairs where the key is the message ID, and the value is the data length. Each time openpilot starts, we don't know what kind of car we have. So we start listening to CAN messages from the car.

For the sake of example, suppose that we receive a message ID 2 with a length of 5 bytes. Given the information provided above, any of the four fingerprints could be valid. Now suppose that we receive a message ID 264 with length 3. Now we can eliminate both fingerprints from CAR.XTRAIL because neither of the fingerprints contains message ID 264. Similarly, if we receive message 42 with length 8, we can eliminate the first CAR.LEAF fingerprint. If, after listening to many more messages, the second CAR.LEAF fingerprint has not been eliminated in this way, we can conclude that the car is a Nissan Leaf.

We use this conclusion to load the correct CAN Dictionaries (DBCs), important information about the geometry and featureset of the car, and functions for reading and writing make/model-specific CAN messages.

Reading the CAN bus is done through selfdrive/car/<make>/carstate and writing to the CAN bus is done through selfdrive/car/<make>/carcontroller.

Every make of car contains a manufacturer-specific, CarInterface, CarState, and CarController, which can be found in the selfdrive/car/<make> directory. When needed, these classes provide differentiation between the vehicle's model.

Now that we have a basic grasp of fingerprinting, let's revisit the first lines of code we looked at. Recall that the line CI.apply(CC) transforms calculations for acceleration and steering angle into make/model specific CAN messages on each loop of the process:

# selfdrive/controls/controlsd.py

# send car controls over can
can_sends = CI.apply(CC)
pm.send('sendcan', can_list_to_can_capnp(can_sends))

The car interface CI is determined by the get_car function, which follows the fingerprinting process of elimination described above.

# selfdrive/controls/controlsd.py

CI, CP = get_car(can_sock, pm.sock['sendcan'])

The get_car function is called once on the initialization of the controlsd process. It passes the CAN Rx topic subscriber, can_sock, and the CAN Tx topic publisher, sendcan, which allows the fingerprint function to send and receive CAN messages.

The goal of the get_car function is to dynamically __import__ the correct CarInterface, CarController, and CarState from the selfdrive/car directories:

# imports from directory selfdrive/car/<make>/
interface_names = _get_interface_names()
interfaces = load_interfaces(interface_names)

def load_interfaces(brand_names):
  ret = {}
  for brand_name in brand_names:
    path = ('selfdrive.car.%s' % brand_name)
    CarInterface = __import__(path + '.interface').CarInterface
    CarState = __import__(path + '.carstate').CarState
    CarController = __import__(path + '.carcontroller').CarController

    for model_name in brand_names[brand_name]:
      ret[model_name] = (CarInterface, CarController, CarState)
  return ret

The candidate from the fingerprinting process is used to select the correct set of CarInterface, CarController, and CarState from the list created by load_interface. A valid candidate is not the key/value dictionary. It is the higher-level key, like CAR.COROLLA, for example:

def get_car(logcan, sendcan):
  candidate, fingerprints, vin, car_fw, source, exact_match = fingerprint(logcan, sendcan)
  ...
  CarInterface, CarController, CarState = interfaces[candidate]
  car_params = CarInterface.get_params(candidate, fingerprints, car_fw)
  ...
  return CarInterface(car_params, CarController, CarState), car_params

The CarInterface.get_params function takes a fingerprint as an input and returns make and model-specific parameters about the car, CP. Here is the get_params function for all Toyotas. Notice how some things like the SafetyModel are the same across all models, while other parameters like the safetyParam and steerRatio differ across models, depending on the candidate:

# selfdrive/car/toyota/interface.py

def get_params(candidate, fingerprint=gen_empty_fingerprint(), car_fw=[]):
    ret = CarInterfaceBase.get_std_params(candidate, fingerprint)

    ret.carName = "toyota"
    ret.safetyModel = car.CarParams.SafetyModel.toyota

    ret.steerActuatorDelay = 0.12  # Default delay, Prius has larger delay
    ret.steerLimitTimer = 0.4

   ...
   elif candidate == CAR.COROLLA:
      stop_and_go = False
      ret.safetyParam = 88
      ret.wheelbase = 2.70
      ret.steerRatio = 18.27
      tire_stiffness_factor = 0.444  # not optimized yet
      ret.mass = 2860. * CV.LB_TO_KG + STD_CARGO_KG  # mean between normal and hybrid
      ret.lateralTuning.pid.kpV, ret.lateralTuning.pid.kiV = [[0.2], [0.05]]
      ret.lateralTuning.pid.kf = 0.00003   # full torque for 20 deg at 80mph means 0.00007818594

    elif candidate == CAR.LEXUS_RX:
      stop_and_go = True
      ret.safetyParam = 73
      ret.wheelbase = 2.79
      ret.steerRatio = 14.8
      tire_stiffness_factor = 0.5533
      ret.mass = 4387. * CV.LB_TO_KG + STD_CARGO_KG
      ret.lateralTuning.pid.kpV, ret.lateralTuning.pid.kiV = [[0.6], [0.05]]
      ret.lateralTuning.pid.kf = 0.00006
   ...

Now that we've loaded the make's interface, and stored the parameters of the model, recall that can_sends = CI.apply(CC) turns calculations for acceleration and steering angle into model-specific CAN messages. But the apply method is just a convenient wrapper for combining the car interface, CI, the car controller, CC, and car state CS through the car controller's update method.

# selfdrive/car/toyota/interface

# pass in a car.CarControl
# to be called @ 100hz
class CarInterface:
  ...
  def apply(self, c):
    can_sends = self.CC.update(c.enabled, self.CS, self.frame, c.actuators, ...)

    self.frame += 1
    return can_sends

The frame property is used to keep track of time, and the actuators argument contains pertinent information about steering and acceleration. Here's the cereal struct definition for actuators:

struct Actuators {
  # range from -1.0 - 1.0
  steer @2: Float32;
  steeringAngleDeg @3: Float32;

  accel @4: Float32; # m/s^2
  longControlState @5: LongControlState;

  enum LongControlState @0xe40f3a917d908282{
   off @0;
   pid @1;
   stopping @2;
   starting @3;
}

The enabled argument determines whether we should override the acceleration and steering, and frame determines how often the message should be sent. The car controllers are quite different depending on the manufacturer and underlying hardware, but here is a simple example of a car controller creating gas and brake CAN messages using the actuators.accel value:

# selfdrive/car/honda/carcontroller.py

class CarController:
  def update(self, enabled, CS, frame, actuators, ...):
    P = self.params
    ...
    if enabled:
      gas, brake = compute_gas_brake(actuators.accel, CS.out.vEgo, CS.CP.carFingerprint)
    else:
      accel = 0.0
      gas, brake = 0.0, 0.0

    can_sends = []

    # apply brake hysteresis
    pre_limit_brake, self.braking, self.brake_steady = actuator_hystereses(brake, self.braking, self.brake_steady, CS.out.vEgo, CS.CP.carFingerprint)

    # wind brake from air resistance decel at high speed
    wind_brake = interp(CS.out.vEgo, [0.0, 2.3, 35.0], [0.001, 0.002, 0.15])

    # send at 50Hz instead of 100Hz
    if (frame % 2) == 0:
      idx = frame // 2
      # decide whether to brake, and remember it
      apply_brake = clip(self.brake_last - wind_brake, 0.0, 1.0)
      apply_brake = int(clip(apply_brake * P.BRAKE_MAX, 0, P.BRAKE_MAX - 1))
      can_sends.append(hondacan.create_brake_command(self.packer, apply_brake,...))
      self.apply_brake_last = apply_brake

      if CS.CP.enableGasInterceptor:
        # way too aggressive at low speed without this
        gas_mult = interp(CS.out.vEgo, [0., 10.], [0.4, 1.0])
        apply_gas = clip(gas_mult * gas, 0., 1.)
        can_sends.append(create_gas_command(self.packer, apply_gas, idx))

    ...

    return can_sends

It's important to understand that the acceleration and steering values require post-processing for different makes and models. For example, Hondas have separate messages for gas and brake, while Toyotas have a single acceleration message.

The job of selfdrive/car/<make>/carcontroller.py is to translate the setpoints into manufacturer specific CAN messages. It is common for the carcontroller to create custom CAN messages using an additional library like selfdrive/car/honda/hondacan.py, which creates manufacturer specific CAN messages using the packer library.

# selfdrive/car/honda/hondacan.py

# CAN bus layout with relay
# 0 = ACC-CAN - radar side
# 1 = F-CAN B - powertrain
# 2 = ACC-CAN - camera side
# 3 = F-CAN A - OBDII port

def get_pt_bus(car_fingerprint):
  return 1 if car_fingerprint in HONDA_BOSCH else 0

def create_brake_command(packer, apply_brake, pump_on, ...):
  brakelights = apply_brake > 0
  brake_rq = apply_brake > 0
  pcm_fault_cmd = False

  values = {
    "COMPUTER_BRAKE": apply_brake,
    "BRAKE_PUMP_REQUEST": pump_on,
    "CRUISE_OVERRIDE": pcm_override,
    "CRUISE_FAULT_CMD": pcm_fault_cmd,
    "CRUISE_CANCEL_CMD": pcm_cancel_cmd,
    "COMPUTER_BRAKE_REQUEST": brake_rq,
    "SET_ME_1": 1,
    "BRAKE_LIGHTS": brakelights,
    ...
  }

  bus = get_pt_bus(car_fingerprint)
  return packer.make_can_msg("BRAKE_COMMAND", bus, values, idx)

The packer function is attached to the car controller and passed as an argument. It is initialized with the name of the DBC file from opendbc, a set of CAN dictionaries.

The job of the packer is to make_can_msg. Here is the DBC file excerpt from the message above.

# opendbc/honda_odyssey_exl_2018_generated.dbc
...
BO_ 506 BRAKE_COMMAND: 8 ADAS
 SG_ COMPUTER_BRAKE : 7|10@0+ (1,0) [0|1] "" EBCM
 SG_ SET_ME_X00 : 13|5@0+ (1,0) [0|1] "" EBCM
 SG_ BRAKE_PUMP_REQUEST : 8|1@0+ (1,0) [0|1] "" EBCM
 SG_ SET_ME_X00_2 : 23|3@0+ (1,0) [0|1] "" EBCM
 SG_ CRUISE_OVERRIDE : 20|1@0+ (1,0) [0|1] "" EBCM
 SG_ SET_ME_X00_3 : 19|1@0+ (1,0) [0|1] "" EBCM
 SG_ CRUISE_FAULT_CMD : 18|1@0+ (1,0) [0|1] "" EBCM
 SG_ CRUISE_CANCEL_CMD : 17|1@0+ (1,0) [0|1] "" EBCM
 SG_ COMPUTER_BRAKE_REQUEST : 16|1@0+ (1,0) [0|1] "" EBCM
...

To learn more about the DBC file protocol, check out @energee's excellent article.