Question

Based on ORB messages, it’s possible to record and replay arbitrary parts of the
system. For this to work, the new logger needs to be enabled (SYS_LOGGER set
to 1).

Replay is useful to test the effect of different parameter values based on real
data, compare different estimators, etc.

Prerequisites

The first thing that needs to be done is to identify the module or modules that
should be replayed. Then, all the inputs to these modules, i.e. subscribed ORB
topics. For system-wide replay, this consists of all hardware input: sensors, RC
input, mavlink commands and file system.

All identified topics need to be logged at full rate (see
logging). For ekf2 this is already the case with the
default set of logged topics.

It is important that all replayed topics contain only a single absolute
timestamp, which is the automatically generated field timestamp. Should there
be more timestamps, then they must be relative with respect to the main
timestamp. For an example, see
sensor_combined.msg.
Reasons for this are given below.

Usage

• First, choose the file to replay, and build the target (from within the
  Firmware directory):

  export replay=<absolute_path_to_log_file.ulg>
  make posix_sitl_default

  This will create the output in a separate build directory
  build/posix_sitl_default_replay (so that the parameters don’t interfere with
  normal builds). It’s possible to choose any posix SITL build target for
  replay, the build system knows through the replay environment variable that
  it’s in replay mode.

• Add ORB publisher rules file in
  build/posix_sitl_default_replay/tmp/rootfs/orb_publisher.rules.
  This file defines which module is allowed to publish which messages. It has
  the following format:

  restrict_topics: <topic1>, <topic2>, ..., <topicN>
  module: <module>
  ignore_others: <true/false>

  It means that the given list of topics should only be published by <module>
  (which is the command name). Publications to any of these topics from another
  module are silently ignored. If ignore_others is true, then publications
  to other topics from <module> are ignored.

  For replay, we only want the replay module to be able to publish the
  previously identified list of opics. So for replaying ekf2, the rules file
  looks like this:

  restrict_topics: sensor_combined, vehicle_gps_position, vehicle_land_detected
  module: replay
  ignore_others: true

  This allows that the modules, which usually publish these topics, don’t need
  to be disabled for replay.

• Optional: setup parameter overrides in the file
  build/posix_sitl_default_replay/tmp/rootfs/replay_params.txt.
  This file should contain a list of <param_name> <value>, like:

  EKF2_GB_NOISE 0.001

  By default, all parameters from the log file are applied. When a parameter
  changed during recording, it will be changed as well at the right time during
  replay. A parameter in the replay_params.txt will override the value and
  changes to it from the log file will not be applied.

• Optional: copy dataman missions file from the SD card to the build
  directory. Only necessary if a mission should be replayed.

• Start the replay:

  make posix_sitl_default jmavsim

  This will automatically open the log file, apply the parameters and start
  to replay. Once done, it will be reported and the process can be exited. Then
  the newly generated log file can be analyzed, it has _replayed appended to
  its file name.

  Note that the above command will show the simulator as well, but depending on
  what is being replayed, it will not show what’s actually going on. It’s
  possible to connect via QGC and e.g. view the changing attitude during replay.

• Finally, unset the environment variable, so that the normal build targets
  are used again:

  unset replay

Important Notes

• During replay, all dropouts in the log file are reported. These have a
  negative effect on replay and thus it should be taken care that dropouts are
  avoided during recording.
• It is currently only possible to replay in ‘real-time’, meaning as fast as the
  recording was done. This is planned to be extended in the future.
• A message that has a timestamp of 0 will be considered invalid and not be
  replayed.

EKF2 Replay

This is a specialization of the system-wide replay for fast EKF2 replay. It will
automatically create the ORB publisher rules and works as following:

• Optionally setSDLOG_MODE to 1 to start logging from boot
• Record the log
• To replay:

export replay_mode=ekf2
export replay=<abs_path_to_log.ulg>
make posix none

You can stop it after there’s an output like:

INFO  [replay] Replay done (published 9917 msgs, 2.136 s)

The parameters can be adjusted as well. They can be extracted from the log with
(install pyulog with sudo pip install pyulog first):

ulog_params -i $replay -d ' ' | grep -e '^EKF2' > build/posix_sitl_default_replay/tmp/rootfs/replay_params.txt

Then edit the parameters in the file as needed and restart the replay process
with make posix none. This will create a new log file.

The location of the generated log is printed with a message like this:

INFO  [logger] Opened log file: rootfs/fs/microsd/log/2017-03-01/13_30_51_replayed.ulg

When finished, use unset replay; unset replay_mode to exit the replay mode.

Behind the Scenes

Replay is split into 3 components:

• a replay module
• ORB publisher rules
• time handling

The replay module reads the log and publishes the messages with the same speed
as they were recorded. A constant offset is added to the timestamp of each
message to match the current system time (this is the reason why all other
timestamps need to be relative). The command replay tryapplyparams is executed
before all other modules are loaded and applies the parameters from the log and
user-set parameters. Then as the last command, replay trystart will again
apply the parameters and start the actual replay. Both commands do nothing if
the environment variable replay is not set.

The ORB publisher rules allow to select which part of the system is replayed, as
described above. They are only compiled for the posix SITL targets.

The time handling is still an open point, and needs to be implemented.