CollectionsCodeDatasets

BenchBot Add-ons Manager

qcr/benchbot_addons

NOTE: this software is part of the BenchBot software stack, and not intended to be run in isolation (although it can be installed independently through pip if desired). For a working BenchBot system, please install the BenchBot software stack by following the instructions here.

BenchBot Add-ons Manager

BenchBot project QUT Centre for Robotics Open Source Primary language License

The BenchBot Add-ons Manager allows you to use BenchBot with a wide array of additional content, and customise your installation to suite your needs. Semantic Scene Understanding not your thing? Install the Semantic Question Answering add-ons instead. Want to create your own content? Write some basic YAML files to make your own add-ons. Need to re-use existing content? Simply include a dependency on that add-on. Add-ons are all about making BenchBot whatever you need it to be—build a BenchBot for your research problems, exactly as you need it.

Add-ons come in a variety of types. Anything that you may need to customise for your own experiments or research, should be customisable through an add-on. If not, let us know, and we'll add more add-on enabled functionality to BenchBot!

The list of currently supported types of add-ons are:

  • batches: a list of environments used for repeatable evaluation scores with the benchbot_batch script.
  • environments: simulated or real world environments that a task can be performed in, with a robot. Only Isaac Sim simulation is currently supported, but there is capacity to support other simulators. Please get in contact if you'd like to see another simulator in BenchBot!
  • evaluation_methods: a method for evaluating a set of formatted results, against a corresponding ground truth, and producing scores describing how well a result performed a given task.
  • formats: formalisation of a format for results or ground truth data, including helper functions.
  • ground_truths: ground truth data in a declared format, about a specific environment. Environments can have many different types of ground truths depending on what different tasks require.
  • robots: a robot definition declaring the communication channels available to the BenchBot ecosystem. Both simulated and real world robots are supported, they just need to run ROS.
  • tasks: a task is a definition of something we want a robot to do, including what observations and actions it has available, and how results should be reported.

See the sections below for details of how to interact with installed add-ons, how to create your own add-ons, and formalisation of what's required in an add-on.

Installing and using the add-ons manager

In general, you won't use the add-ons manager directly. Instead you interact with the BenchBot software stack, which uses the add-ons manager to manage and access add-ons.

The manager is a Python package if you do find you want to use it directly, and installable with pip. Run the following in the root directory where the repository was cloned:

[email protected]:~$ pip install .

The manager can then be imported and used to manage installation, loading, accessing, processing, and updating of add-ons. Some samples of supported functionality are shown below:

from benchbot_addons import manager as bam

# Check if example with 'name' = 'hello_scd' exists
bam.exists('examples', [('name', 'hello_scd')])

# Find all installed environments
bam.find_all('environments')

# Get a list of the names for all installed tasks
bam.get_field('tasks', 'name')

# Get a list of (name, variant) pairs for all installed environments
bam.get_fields('environments', ['name', 'variant'])

# Find a robot with 'name' = 'carter'
bam.get_match('robots', [('name', 'carter')])

# Get the 'results_format' value for the task called 'scd:passive:ground_truth'
bam.get_value_by_name('tasks', 'scd:passive:ground_truth', 'results_format')

# Load YAML data for all installed ground truths
bam.load_yaml_list(bam.find_all('ground_truths', extension='json'))

# Install a list of comma-separated add-ons
bam.install_addons('benchbot-addons/ssu,benchbot-addons/sqa')

# Install a specific add-on (& it's dependencies)
bam.install_addon('tasks_ssu')

# Print the list of currently installed add-ons, & officially available add-ons
bam.print_state()

# Uninstall all add-ons
bam.remove_addons()

# Uninstall a string separated list of add-ons
bam.remove_addon('benchbot-addons/ssu,benchbot-addons/sqa')

Creating your own add-on content

Add-ons are designed to make it easy to add your own local content to a BenchBot installation. You can add your own local content to the "local add-ons" folder provided with your install. The location on your machine can be printed via the following:

from benchbot_addons import manager as bam

print(bam.local_addon_path())

BenchBot expects add-on content to be in named folders denoting the type of content. For example, robots must be in a folder called 'robots', tasks in a folder called 'tasks', and so on. A list of valid content types is available via the SUPPORTED_TYPES field in the add-ons manager.

Below is an example of the process you would go through to create your own custom task locally:

  1. Find the location for your custom local add-ons:
    [email protected]:~$ python3 -c 'from benchbot_addons import manager as bam; print(bam.local_addon_path())'
    /home/ben/repos/benchbot/addons/benchbot_addons/.local/my_addons
    
  2. Create the following YAML file for your task: /home/ben/repos/benchbot/addons/benchbot_addons/.local/my_addons/tasks/my_task.yaml
  3. Use the fields described below in the task add-ons specification to define your task
  4. Save the file

Done. Your new custom task should now be available for use in your BenchBot system (e.g. benchbot_run --list-tasks).

Sharing your custom add-ons

Custom add-on content can be grouped together into an add-on package, of which there are two different types: 'official' and third-party.

'Official' packages are those we've verified, and are stored in our benchbot-addons GitHub organisation. You can get a full list of official add-on packages through the manager.official_addons() helper function, or benchbot_install --list-addons script in the BenchBot software stack.

Third-party add-on packages differ only in that we haven't looked at them, and they can be hosted anywhere on GitHub you please.

Creating all add-on packages is exactly the same process, the only difference is whether the repository is inside or outside of the benchbot-addons GitHub organisation:

  1. Create a new GitHub repository
  2. Add folders corresponding to the type of content your add-ons provide (i.e. an environments add-on has an environments directory at the root).
  3. Add YAML / JSON files for your content, and make sure they match the corresponding format specification from the section below
  4. Add in any extra content your add-on may require: Python files, simulator binaries, images, etc. (if your add-on gets too big for a Git repository, you can zip the content up, host it somewhere, and use the .remote metadata file described in the next section)
  5. Decide if your package has any dependencies, and declare them using the appropriate .dependencies* files
  6. Push everything up to GitHub on your default branch

Note: it's a good idea to only include one type of add-on per repository as it makes your add-on package more usable for others. It's not a hard rule though, so feel free to add multiple folders to your add-on if you require.

Feel free to have a look at any of the official add-ons for help and examples of how to work with add-ons.

Add-ons format specification

Here are the technical details of what's expected in add-on content. The BenchBot system will assume these specifications are adhered to, and errors can be expected if you try to use add-ons that don't match the specifications.

An add-on package has the following structure (technically none of the files are required, they just determine what functionality your add-on includes):

Filename Description
.dependencies A list of add-on packages that must be installed with this package. Packages are specified by their GitHub identifier (i.e. github_username/repository_name), with one per line
.dependencies-python A list of Python dependencies for your add-on. Syntax for file is exactly the same as requirements.txt files.
.remote Specifies content that should be installed from a remote URL, rather than residing in this repository. A remote resource is specified as a URL and target directory separated by a space. One resource is specified per line. The add-ons manager will fetch the URL specified, and extract the contents to the target directory (e.g. http://myhost/my_content.zip environments)
<directory>/ Each named directory corresponds to an add-on type described below. The directory will be ignored if its name doesn't exactly match any of those below.

Batch add-ons

A YAML file, that must exist in a folder called batches in the root of the add-on package (e.g. batches/my_batch.yaml).

The following keys are supported for batch add-ons:

Key Required Description
'name' Yes A string used to refer to this batch (must be unique!).
'environments' Yes A list of environment strings of the format 'name':'variant' (e.g. 'miniroom:1').

Environment add-ons

A YAML file, that must exist in a folder called environments in the root of the add-on package (e.g. environments/my_environment.yaml).

The following keys are supported for environment add-ons:

Key Required Description
'name' Yes A string used to refer to this environment's name (the ('name', 'variant') pair must be unique!).
'variant' Yes A string used to refer to this environment's variant (the ('name', 'variant') pair must be unique!).
'type' Yes A string describing the type of this environment ('sim_unreal' & 'real' are the only values currently used).
'map_path' Yes A path to the map for this environment, which will be used by either the simulator or real world system to load the environment.
'start_pose' Yes The start pose of the robot that will be provided to users through the BenchBot API. The pose is specified as a list of 7 numbers: quarternion_w, quarternion_x, quarternion_y, quarternion_z, position_x, position_y, position_z. This must be accurate!
'trajectory_poses' No A list of poses for the robot to traverse through in order. Each pose is a list of 7 numbers: quarternion_w, quarternion_x, quarternion_y, quarternion_z, position_x, position_y, position_z. This environment won't be usable for tasks that use the 'move_next' action if this parameter isn't provided.
'robots' No A list of supported names for robot that are supported in this environment. If this list isn't included, all robots with the same 'type' as this environment will be able to run.
'object_labels' No A list of labels for the objects that exist in the scene. Can be used with simulated sensors like segmentation sensors.

Evaluation method add-ons

A YAML file, that must exist in a folder called evaluation_methods in the root of the add-on package (e.g. evaluation_methods/my_evaluation_method.yaml).

The following keys are supported for evaluation method add-ons:

Key Required Description
'name' Yes A string used to refer to this evaluation method (must be unique!)
'valid_result_formats' Yes List of strings denoting results formats supported by the evaluation method. Ideally these format definitions should also be installed.
'valid_ground_truth_formats' Yes List of strings denoting ground truth formats supported by the evaluation method. Ideally these format definitions should also be installed.
'functions' Yes Dictionary of named functions provided by the evaluation method. The named methods are key value pairs where the key is the function name, and the value is a string describing how the function can be imported with Python. For example, evaluate: "omq.evaluate_method" declares a function called 'evaluate' that is imported via from omq import evaluate_method. Likewise "omq.submodule.combine_method" translates to from omq.submodule import combine_method. See below for the list of functions expected for evaluation methods.
'description' No A string describing what the evaluation method is and how it works. Should be included if you want users to understand where your method can be used.

Evaluation methods expect the following named functions:

Name Signature Usage
'evaluate' fn(dict: results, list: ground_truths) -> dict Evaluates the performance using a results dictionary, and returns a dictionary of containing the scores. It also takes a list of dictionaries containing each ground truth that will be used in evaluation.
'combine' fn(list: scores) -> dict Takes a list of scores dictionaries, and returns an aggregate score. If this method isn't declared, benchbot_eval won't return a summary score.

Example method add-ons

A YAML file, that must in a folder called examples in the root of the add-on package (e.g. examples/my_example.yaml).

The following keys are supported for example add-ons:

Key Required Description
name Yes A string used to refer to this example (must be unique!)
native_command Yes A string describing the command used to run your example natively, relative to the directory of this YAML file! For example running your my_example.py file which is in the same director as this YAML would be python3 ./my_example.py.
container_directory No Directory to be used for Docker's build context. The submission process will automatically look for a file called Dockerfile in that directory unless the 'container_filename' key is also provided.
container_filename No Custom filename for your example's Dockerfile. Dockerfile in container_directory will be used if this key is not included. This path is relative to this YAML file, not 'container_directory'.
description No A string describing what the example is and how it works. Should be included if you want users to understand how your example can be expanded.

Format definition add-ons

A YAML file, that must exist in a folder called formats in the root of the add-on package (e.g. formats/my_format.yaml).

The following keys are supported for format add-ons:

Key Required Description
'name' Yes A string used to refer to this format (must be unique!)
'functions' Yes Dictionary of named functions for use with this format. The named methods are key-value pairs where the key is the function name, and the value is a string describing how the function can be imported with Python. For example, create: "object_map.create_empty" declares a function called 'create' that is imported via from object_map import create_empty. Likewise "object_map.submodule.validate" translates to from object_map.submodule import validate. See below for the list of functions expected for format definitions.
'description' No A string describing what the format is and how it works. Should be included if you want users to understand what your format is supposed to capture.

Format definitions expect the following named functions:

Name Signature Usage
'create' fn() -> dict Function that returns an empty instance of this format. As much as possible should be filled in to make it easy for users to create valid instances (especially when a format is used for results).
'validate' fn(dict: instance) -> None Takes a proposed instance of this format and validates whether it meets the requirements. Will typically use a series of assert statements to confirm fields are valid.

Ground truth add-ons

A JSON file, that must exist in a folder called ground_truths in the root of the add-on package (e.g. ground_truths/my_ground_truth.json).

The following keys are supported for ground truth add-ons:

Key Required Description
'environment' Yes A dictionary containing the definition data for the ground truth's reference environment. The data in this field should be a direct copy of an environment add-on.
'format' Yes A dictionary containing the definition data for the ground truth's format. The data in this field should be a direct copy of a format definition add-on.
'ground_truth' Yes A valid instance of the format described by the 'format' field. This is where your actual ground truth data should be stored.

A lot of these keys should be copied from other valid definitions. Please see the GroundTruthCreator helper class in BenchBot Evaluation for assistance in creating valid ground truths.

Robot add-ons

A YAML file, that must exist in a folder called robots in the root of the add-on package (e.g. robots/my_robot.yaml).

The following keys are supported for robot add-ons:

Key Required Description
'name' Yes A string used to refer to this robot (must be unique!).
'type' Yes A string describing the type of this robot ('sim_unreal' & 'real' are the only values currently used).
'address' Yes A string for the address where a running BenchBot Robot Controller can be accessed (e.g. 'localhost:10000')
'global_frame' Yes The name of the global TF frame. All poses reported by the BenchBot API will be with respect to this frame.
'robot_frame' Yes The name of the robot's TF frame.
'poses' Yes A list of named poses that this robot provides. This list of poses will be available in observations provided by the BenchBot API.
persistent_cmds Yes A list of commands that will be run and kept alive for the lifetime of the robot controller. The commands will be run in parallel, and executed via bash -c <your_command_string>
persistent_status Yes A command used to check the status of your persistent_cmds. This command should execute quickly, and terminate on completion, with the return code being used to evaluate the status. The command string is executed via bash -c <your_command_string>
run_cmd Yes A single command issued by the controller to run a simulation. This command must terminate on completion. The command string is executed via bash -c <your_command_string>
stop_cmd Yes A single command issued by the controller to stop a simulation. This command must terminate on completion. The command string is executed via bash -c <your_command_string>
'connections' Yes A dictionary of connections that your robot makes available to the BenchBot ecosystem. The name of the key-value pair is important, and should follow the recommendations provided on standard channels in the BenchBot API documentation. A description of connection definitions is provided below.

Connections are the lifeblood of interaction between BenchBot and robot platforms. They are defined by named entries, with the following fields:

Key Required Description
'connection' Yes Connection type string, used by the BenchBot Robot Controller. Supported values are 'api_to_ros' (used for actions), 'ros_to_api' (used for observations), and 'roscache_to_api' (special value used for caching observation values).
'ros_topic' Yes Topic name for the ROS side of the connection.
'ros_type' Yes Topic type for the ROS side of the connection.
'callback_api' No A callback that is run on the HTTP encoded data received / sent on the API end of the connection. It takes in data, and returns transformed data based on the callback's action. Callbacks are specified by a string denoting how the callback can be accessed (e.g. 'api_callbacks.convert_to_rgb = from api_callbacks import convert_to_rgb). No data transformation occurs if no callback is provided.
'callback_ros' No A callback that is run on the ROS data received / sent on the robot controller end of the connection. It takes in data and a reference to the robot controller. 'api_to_ros' connections use this data to act on the robot, whereas 'ros_to_api' connections turn this data into a dictionary that can be serialised into HTTP traffic. Callbacks are specified by a string denoting how the callback can be accessed (e.g. 'api_callbacks.convert_to_rgb = from api_callbacks import convert_to_rgb). No action occurs at the ROS level if no callback is provided.

Task add-ons

A YAML file, that must exist in a folder called tasks in the root of the add-on package (e.g. tasks/my_task.yaml).

The following keys are supported for task add-ons:

Key Required Description
'name' Yes A string used to refer to this task (must be unique!).
'actions' Yes A list of named connections to be provided as actions through the BenchBot API. Running this task will fail if the robot doesn't provide these named connections.
'observations' Yes A list of named connections to be provided as observations through the BenchBot API. Running this task will fail if the robot doesn't provide these named connections.
'localisation' No A string describing the level of localisation. Only supported values currently are 'ground_truth' and 'noisy'. The default value is 'ground_truth'.
'results_format' No A string naming the format for results. The format must be installed, as BenchBot API will use the format's functions to provide the user with empty results.
'description' No A string describing what the task is, and how it works. Should be included if you want users to understand what challenges your task is trying to capture.
'type' No A string describing what robot / environment types are valid for this task. For example, a task that provides a magic image segmentation sensor would only be made available for 'sim_unreal' type robots / environments.
'scene_count' No Integer representing the number of scenes (i.e. environment variations required for a task). If omitted, a default value of 1 will be used for the task.

CRICOS No. 00213J