Skip to content

Benchmarking Platform for Classic and Learned Path Planning Algorithms.

License

Notifications You must be signed in to change notification settings

perfectly-balanced/PathBench

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

PathBench: Benchmarking Platform for Classic and Learned Path Planning Algorithms

coverage report

PathBench is a motion planning platform used to develop, assess, compare and visualise the performance and behaviour of both classic and machine learning-based robot path planners in a two- or three-dimensional space.

Quick Start

Python 3.8.5 officially supported, though older (and newer) versions should work for the most part.

The following installation and run instructions have been used for running PathBench on Ubuntu 18 and 20.

Installing dependencies

pip3 install -r requirements.txt

Optional dependency is ompl with installation not covered here, although installation is detected automatically if both ompl and its Python bindings are installed. There are some extra dependencies needed for testing, and are detailed here.

Simulator Visualiser Usage

python3 src/main.py -v

Note, the main script can be run from any working directory as PathBench does not use relative paths internally, however this document will specify commands as if run from the root directory for clarity.

Key Action
escape Exit the simulator
mouse left click Moves agent to mouse location
mouse right click Moves goal to mouse location
arrow keys, PgUp, PgDn Move agent (goal when Alt down) along the X-Y-Z plane
w a s d Rotate orbital camera around map
t Find the path between the agent and goal
x Pause/Resume path finding (animations required)
p Take screenshot
o Take top-view high resolution screenshot of the map
c Toggle visibility of simulator config window
v Toggle visibility of view editor window
m Toggle map between Sparse and Dense
i Toggle visibility of debug overlay

Note, screenshots are placed in data/screenshots/.

Other components

Example usage:

# Run trainer
python3 src/run_trainer.py --num_maps 10000 --full-train --model LSTM
# Run generator
python3 src/main.py --generator --num-maps 100 --generator-type house --dims 3
# Run analyzer
python3 src/main.py --analyzer --algorithms "A*" Dijkstra VIN --include-all-builtin-maps

Note A* is in quotes to prevent glob expansion by the shell.

You can specify multiple components to be run in order: for example, running the trainer and then visualising the resulting trained ML algorithm:

python3 src/main.py --trainer --visualiser

More Options

Run

python3 src/main.py --help

to get a full list of all available options for each of the generator, analyser, trainer, and visualiser.

Testing

Dependencies

pip3 install -r tests/requirements.txt
sudo apt-get install scrot

Headless Testing

Dependencies

sudo apt-get install x11vnc xvfb xtightvncviewer

Example Usage

  1. Running an individual graphics test (labyrinth with A* in this case):
python3 tests/test_graphics/test_labyrinth_A.py --spawn-display --view-display
  1. Running all tests:
python3 tests/run_tests.py --spawn-display --view-display
  • Specifying --view-display will internally execute vncviewer. This will result in a white dialog popup. Press Enter when it appears, and the interactive view will subsequently appear. However, please wait approximately a second before pressing Enter, otherwise the viewer will be created before the view server has had time to initialise, which will prevent launching the interactive view (i.e. no-op).
  • Specifying --spawn-display for run_tests.py will launch an isolated virtual display for each test as they have side-effects, which would cause failures on the CI. As a result, when specifying --view-display, the viewer will keep relaunching itself for each graphics test.
  • For more usage details specify the help flag -h.

Note, to view a screenshot of the screen, execute the following:

xwud -in /var/tmp/Xvfb_screen0

PathBench

Simulator

This section is responsible for environment interactions and algorithm visualisation. It provides custom collision detection systems and a graphics framework (Panda3D) for rendering the internal state of the algorithms in 2D or 3D. In PathBench, a simulation can be run headlessly or with graphics.

The GUI main window consists of several components - a Simulator Configuration, View Editor and a Debug Overlay.

PathBench Main Window

The Simulator Configuration window is used to make a selection between different maps and algorithms to analyse, set goal and start positions and change animation settings. A map is then initialised on pressing "Update".

The View Editor window is used to customise the map, e.g. change colours of entities or set transparency level. Those modifications could also be saved as one of the six states using the "Save" button, and the changes can be reverted at any point using "Restore". The 1-6 state buttons provide the freedom to easily toggle between different custom states.

The Debug Overlay at the top left of the main window provides information about the selected map and algorithm, the entity positions, and the current state of the simulator (Initialising, Initialised, Running, Done, Terminated).

To run and use the simulator see Quick Start.


Generator

This section is responsible for generating and labelling the training data used to train the Machine Learning models.

It has several options to change the generation algorithm and vary hyperparameters to that algorithm, as well as specifying number of dimensions and size of the maps. They are saved to a directory corresponding to data/maps/[algorithm]_[number](_3D)/[0..number].json, and can then be used for training, analysing, or for use in the visualiser.

Trainer

This section is a class wrapper over the third party Machine Learning libraries. It provides a generic training pipeline based on the holdout method and standardised access to the training data. It is a wrapper over the Pytorch python package for machine learning and provides a generic training pipeline for path planning models based on the holdout method and standardized access to the training data generated by the Generator.

Each learned model in PathBench3D (LSTM, CAE, Combined LSTM+CAE and also VIN) inherits from the MLModel class, which trains its models using a standard holdout method. The models are trained on the data labels generated by the Generator class, using data received from the A* algorithm. Each model contains metadata about what labels it needs. For example, LSTM trains on raycast information and angle to the goal, whereas VIN just uses the map and path.

Analyzer

The final section manages the statistical measures used in the practical assessment of the algorithms. Custom metrics can be defined as well as graphical displays for visual interpretations.

The current metrics are:

  • Average Path Deviation
  • Success Rate
  • Average Time
  • Average Steps
  • Average Distance
  • Average Distance from Goal
  • Average Original Distance from Goal
  • Average Trajectory Smoothness
  • Average Obstacle Clearance
  • Average Search Space
  • Maximum Memory

These are saved to pbtest.csv, and plotted on bar charts and violin plots using matplotlib as below:

ROS Real-time Extension

This extension provides real-time support for visualisation, coordination and interaction with a physical robot.

  • For a basic demonstration of interacting with RosMap see here.
  • For the fully functional 2D ROS Real-time Extension see here.

Example Trajectory

This shows a screenshot of the ROS extension running with Gazebo emulating a real robot (turtlebot3).

Architecture High Overiew

Platform Architecture

Infrastructure

The MainRunner component is the main entry point of the platform and it coordinates all other sections. The MainRunner takes a master Configuration component as input which represents the main inflexion point of the platform. It describes which section(s) (Simulator, Generator, Trainer, Analyser) should be used and how.

The Services component is a bag of Service components which is injected into all platform classes in order to maintain global access to the core libraries. A Service component is created for most external libraries to encapsulate their APIs and provide useful helper functions. Moreover, by making use of the Adapter Pattern we can easily switch third party libraries, if needed, and the code becomes more test friendly. Finally, the Services container can be mocked together with all its Service components, thus avoiding rendering, file writing and useless printing.

The Simulator was build by following the Model-View-Controller (MVC) pattern. The Model represents the logic part, the View renders the Model and the Controller handles the input from the keyboard and mouse, and calls the appropriate functions from the associated Model.

The EventManager component is a communication service which allows the Model to update the View as there is no direct connection between them (from Model to View, the other way is).

The Debug component is a printing service which augments printing messages with different decorators such as time-stamp and routes the messages to a specified IO stream or standard out. It also provides a range of debugging/printing modes: None (no information), Basic (only basic information), Low (somewhat verbose), Medium (quite verbose), High (all information).

The Torch service is not an actual wrapper around pytorch, but instead it defines some constants such as the initial random seed and the training device (CPU/CUDA).

The Resources service is the persistent storage system. It is a container of Directory components which represent an interface over the actual filesystem directories. It provides safe interaction with the filesystem and a range of utility directories: Cache (temporary storage used for speeding up second runs), Screenshots, Maps (stores all user defined and generated maps), Images (stores images which can be converted to internal 2D maps), Algorithms (stores trained machine learning models), Training Data (stores training data for machine learning models). The main serialisation tool is dill which is a wrapper around pickle with lambda serialisation capabilities, but JSON support has been added, and this should be preferred when possible. Custom serialisation is also allowed such as tensor serialisation provided by pytorch.

The AlgorithmRunner manages the algorithm session which contains the Algorithm, BasicTesting and Map. The AlgorithmRunner launches a separate daemon thread that is controlled by a condition variable. When writing an Algorithm, special key frames can be defined (e.g. when the trace is produced) to create animations. Key frames release the synchronisation variable for a brief period and then acquire it again, thus querying new rendering jobs. Animation is done automatically by monitoring of the trace and display components returned by set_display_info(), with the visualiser updated during the keyframe call.

The Utilities section provides a series of helper methods and classes: Maps (holds in-memory user defined Map components), Point, Size, Progress (progress bar), Timer, MapProcessing (feature extractor used mainly in ML sections).