Oracle Motion Cueing Algorithm#

The demo_oracle_predictor application demonstrates the integration of an "Oracle" predictor and the controller using the Pipeline. That is, the pipeline namespace contains a number of InOutput and Step classes that, when used together:

  1. can be used to calculate a prediction (output reference),
  2. communicate that to the controller,
  3. retrieve the control commands to be sent to the simulator, and
  4. deal with potential computational overruns of the iterative solver in the controller.

Some of the InOutput and Step types used in this example are class templates taking the simulator class as argument. To use these InOutput and Step types, they need to be instantiated explicitly somewhere in the program using them. This page documents how.

This example uses the MpmcaHexapod simulator and thus links against the mpmca_hexapod::mpmca_hexapod target, see the CMakeLists.txt in the application folder.

Requires a dataset download

This example requires you to download data from an external source. The required datasets are quite large, so you might want to start the download first. See the section below.

Configuration files

In this example, the Pipeline is configured using configuration files, see this tutorial.

Message bus

In this example, the InOutput and Step instances communicate with each other through the MessageBus, see this tutorial.

Foxglove Studio

The data generated in this example can be visualized using Foxglove Studio. See this page for instructions on how to install and set it up.

Download dataset#

Several free and open-source car driving datasets related to Autonomous Driving applications exist that can be used to get started with the MPMCA if you do not have your own data. Typically, these datasets contain measurements from an Inertial Measurement Unit (IMU), providing linear accelerations or specific forces and rotational rates (6 signals in total). Note that the MPMCA takes 9 signals as output reference: specific forces, rotational rates, and rotational accelerations. If you have a dataset without rotational accelerations, you can simply provide zeros for the reference rotational accelerations and specify an output error weight of 0 for those three channels.

In this example, we use the Audi Autonomous Driving Dataset (A2D2).1 The A2D2 dataset is very large, but you do not have to download it completely to follow this example. You can either use a simple script provided in the data folder or download the data yourself.

Download script#

Run the download_unpack_data.sh bash script inside the data/a2d2 folder to automatically download and unpack the Gaimersheim data file to the correct location. To run the script, move into the data/a2d2 folder and give the script execution rights, by issuing the following command:

chmod +x download_unpack_data.sh

Then, run the script through

./download_unpack_data.sh

You can use the config.json file as provided in the applications/demo_oracle_predictor directory.

Download data manually#

From the download page, download one of the "Bus signals" datasets (Gaimersheim, Ingolstadt, or Munich), or download the "Dataset Preview", which also contains a small "Bus signals" dataset. This example assumes you to use the "Gaimersheim" file. Extract the file called bus_signals_20180810150607.json from the camera_lidar/20180810_150607/ folder in the archive, and place it in the data/a2d2/ folder inside this repository. If you use a different file or place it on a different location, make sure you adjust the config.json file inside applications/demo_oracle_predictor.

Steps and InOutputs#

This section documents the Step and InOutputs involved in this example.

Required includes#

The main.cpp of this example application needs to include the following files to be able to instantiate the InOutput and Step types that are not templated:

1#include "mpmca/pipeline/in_output/go_to_run.hpp"
2#include "mpmca/pipeline/step/json_playback.hpp"

The other InOutput and Step types are linked into the program in the other .cpp files in the application folder. These InOutput and Step types are discussed below.

ControlCommands#

The file control_commands.cpp instantiates a ControlCommands<TSIM> type for the MpmcaHexapod simulator, and registers it with the InOutputRegister (to make sure it can be instantiated through a configuration file).

 1#include "mpmca/pipeline/in_output/control_commands.hpp"
 2
 3#include "mpmca/pipeline/in_output/control_commands.tpp"
 4#include "mpmca_hexapod/mpmca_hexapod.hpp"
 5
 6using namespace mpmca::pipeline;
 7
 8template class in_output::ControlCommands<mpmca::control::MpmcaHexapod>;
 9
10static InOutputRegistrar<
11    in_output::ControlCommands<mpmca::control::MpmcaHexapod>>
12    ControlCommandsRegistrar("ControlCommands");

The tasks of ControlCommands are:

  1. to deal with computational overruns of the PredictiveControl step in the Task,
  2. to provide the correct initial state for each Controller update in MainTick(), and
  3. to provide the correct position-velocity-acceleration (PVA) set-points to be sent to the simulator (from another InOutput).

Computational overruns

Note that the generic functionality related to dealing with computational overruns is implemented inside the Pipeline class. The ControlCommands class implements the more specific behavior necessary to provide the correct control commands to the simulator when computational overruns occur.

FadeInertialReference#

The file fade_inertial_reference.cpp instantiates a FadeInertialReference<TSIM> type for the MpmcaHexapod simulator, and registers it with the StepRegister (to make sure it can be instantiated through a configuration file), in a similar way as was done for ControlCommands.

The tasks of FadeInertialReference are:

  1. to 'fade in' the output reference signals when entering the RUN state,
  2. to 'fade out' the output reference signals when leaving the RUN state, and
  3. to keep the global StateMachine in STOPPING until the simulator has reached zero velocity.

PredictiveControl#

The file predictive_control.cpp instantiates a PredictiveControl<TSIM> type for the MpmcaHexapod simulator, and registers it with the StepRegister, in a similar way as was done for ControlCommands.

Include both the .hpp and .tpp files.

Note that the (templated) implementation of PredictiveControl is placed outside the header file (.hpp) in a separate file with .tpp extension. Both need to be included.

The tasks of PredictiveControl are:

  1. to create an Controller object of the correct type,
  2. to pass all configuration options to the Controller,
  3. to pass all relevant (input) messages from the MessageBus to the Controller,
  4. to call the Prepare and Feedback functions of the Controller in the right order and at the right moment, and
  5. to put relevant data from the Controller onto (output) messages on the MessageBus.

Note that the PredictiveControl step is intended to perform the above listed tasks only, and you should not expect it to do anything else. All other functionality should be implemented in other Step types that receive information from PredictiveControl through messages on the MessageBus. For example, the following tasks are to be handled by other Step types:

  1. to check the validity of all input and output messages on the MessageBus (not part of the code base, to be implemented by the user),
  2. to fade in or out the reference signals provided to PredictiveControl (see FadeInertialReference),
  3. to gracefully start or stop the motion of the actual simulator (also see FadeInertialReference), or
  4. to log information about the state of the Controller to a file on disk (see McapLogger).

JsonPlayback#

The JsonPlayback step does not depend on the type of simulator to be controlled, and thus does not require a separate .cpp file in which a class template is instantiated.

The tasks of JsonPlayback are:

  1. to load time-series data from disk containing a recording of inertial signals,
  2. to construct a valid output reference using the loaded data, based on the current time and the time steps on the prediction horizon.

McapLogger#

The file mcap_logger.cpp instantiates an McapLogger<TSIM> type for the MpmcaHexapod simulator, and registers it with the StepRegister, in a similar way as was done for ControlCommands.

The task of McapLogger is to take relevant data from the MessageBus and save it to an MCAP file, for visualization by Foxglove Studio, see this page.

Run the demo#

To run the demo, execute the following command from the root of the repository:

./build/bin/demo_oracle_predictor -c applications/demo_oracle_predictor/config.json

This will run the Pipeline in 'realtime' mode. That is, the program will try to run exactly as fast as the real world time (the time you see on a wall clock). For the "Dataset preview" data, this will take approximately 8 minutes. The realtime mode is the mode you need when controlling a simulator during a human-in-the-loop experiment. However, when calculating results for later inspection using Foxglove, you might want to let the program run as fast as possible for your hardware. To do so, add the --free arguments:

./build/bin/demo_oracle_predictor -c applications/demo_oracle_predictor/config.json --free

The recorded MCAP data file will be called demo_oracle_predictor.mcap and you can find it in the directory from which you run the application (in this case, the root of the repository).

Results#

The results can be inspected using Foxglove Studio. An example screenshot of typical cueing results is shown below. Click here to enlarge.

Foxglove Studio screenshot.

From the results, observe the following:

  1. The 'phase shift' between the reference output and expected output of the lateral and longitudinal motion cues is very small. That is, the shape of the expected output follows the shape of the reference output precisely, but with some scaling errors.
  2. Very few false cues can be observed in the lateral and longitudinal channels.
  3. The roll and pitch motion is used to provide lateral and longitudinal specific forces (typically referred to as "g-tilt"), but note that this effect was not 'programmed into' the algorithm manually. It just results automatically from the taken approach.
  4. Given the rather limited yaw motion space of a hexapod, the MPMCA will extensively 'pre-position' the simulator to obtain more space. Whether or not the person inside the simulator will be able to feel this pre-positioning depends on many factors. For this particular example, it is likely that it can be felt, as the pre-positioning rotational rate exceeds 5 deg/s.
  5. The MPMCA accurately replicates the 'low frequency' motions, such as cornering accelerations and braking/acceleration, but also replicates the 'high frequency' motions, such as road noise and bumps in the road.

Cueing improvements#

The cueing results obtained with this example will certainly not be the best results possible for this dataset and this simulator. Several improvements can be obtained from the following:

  1. Add the ScaleInertialReference Step to linearly scale the reference output signals. For the MpmcaHexapod simulator, a good starting point is to scale the longitudinal and lateral specific forces by a factor of 0.5, and the yaw rotational rates by 0.4.
  2. Modify the output error weights to emphasize certain degrees-of-freedom over others. In this example, the "magic-10" output error weights are used (i.e., 1.0 for specific forces, and 10.0 for rotational rates), but other values might lead to better results.

Bibliography#


  1. Geyer, J., Kassahun, Y., Mahmudi, M., Ricou, X., Durgesh, R., Chung, A. S., Hauswald, L., Pham, V. H., Mühlegg, M., Dorn, S., Fernandez, T., Jänicke, M., Mirashi, S., Savani, C., Sturm, M., Vorobiov, O., Oelker, M., Garreis, S., & Schuberth, P. (2020). A2D2: Audi Autonomous Driving Dataset. https://www.a2d2.audi