SocNavBench Baseline Algorithms

Welcome to a small repository of baseline algorithm implementations for SocNavBench

This collection contains the implementation for the following robot planner algorithms interfaced through the JoystickAPI to the SocNavBench simulator.

• Random (naive) (python3)
• RandomCpp (naive) (c++)
• Sampling (naive) (python3)
• RVO (c++)
• RVOwCkpt (c++)
• Social forces (c++)

The following require installing the implementation from this repo (see below in SACADRL implementation)

• SACADRL (python3)
• SACADRLwCkpt (python3)

Building instructions

TODO: see todo #2

All c++ implementations should have valid makefiles where you can just run make. This has been tested on Ubuntu 21.04.

NOTE For running the social_force/social_force executable, you'll first need to run

cd /path/to/SocNavBench/joystick/social_force/
export LD_LIBRARY_PATH=./src
# then you can safely run
./social_force

This is not the case for the RVO2 binary which can simply be executed with

cd /path/to/SocNavBench/joystick/RVO2/
./RVO2

NOTE these binaries require executable permissions, so after compilation you'll need to give these with chmod a+x.

SocNavBench integration

first off, you'll need to copy the contents of this entire repo and replace the current joystick/ implementation in SocNavBench

cd /path/to/SocNavBench-baselines/
rm -rf /path/to/SocNavBench/joystick/* # clear old directory
cp -r * /path/to/SocNavBench/joystick/ # copy all new files over

Running instructions

IMPORTANT make sure you already integrated this codebase with SocNavBench as described in SocNavBench integration

For all except the randomCpp joystick implementations, you'll be able to run the python joystickAPI interface through the joystick_client.py file. Eg.

cd /path/to/SocNavBench/
PYTHONPATH='.' python3 joystick/joystick_client.py --algo sampling # sampling planner

You may also be interested in passing specific algorithms to run with the --algo flag. See the argparser details for all available options.

Note that most of the c++ joystick implementations interface through python (the exception being joystick_client.cpp (random c++ joystick)) but will still have separate executables from their respective source code.

• Therefore in order to run the RVO, RVOwCkpt, or SocialForces planners, you'll need three terminals to run:

1. the SocNavBench simulator itself
2. the joystick_client.py client
3. the c++ binary (RVO2 or social_force)

SACADRL implementation

The implementation for SACADRL uses this source code and can be installed as follows:

• Install the dependencies mentioned here
• Clone the repo and copy the gym_collision_avoidance/ directory to joystick/ directly (NOT joystick/sacadrl)
• Then you should be able to run the joystick_client.py script with --algo sacadrl and --also sacadrlwckpt without issue. (Make sure to keep the PYTHONPATH set to the socnav base dir)

TODOs:

• Add unit tests
• Refactor c++ implementations and building instructions
• Add full support for both modes of robot motion (system dynamics or positional dynamics)
  • currently positional dynamics is 100% supported, but system (velocity control) dynamics is still in development