quadruped-reactive-walking

Implementation of a reactive walking controller for quadruped robots. Architecture mainly in Python with some parts in C++ with bindings to Python.

Dependencies

• Install a Python 3 version (for instance Python 3.6)

• For the following installations each time you see something like "py36", replace it by your Python version ("py35" for instance)

• Install Pinocchio: https://stack-of-tasks.github.io/pinocchio/download.html

• Install Gepetto Viewer: sudo apt install robotpkg-py36-qt4-gepetto-viewer-corba

• Install robot data: sudo apt install robotpkg-example-robot-data

• Install Scipy, Numpy, Matplotlib, IPython: python3.6 -m pip install --user numpy scipy matplotlib ipython

• Install PyBullet: pip3 install --user pybullet

• Install OSQP solver: [https://osqp.org/docs/get_started/sources.html#build-the-binaries]

  • git clone --recursive https://github.com/oxfordcontrol/osqp
  • cd osqp
  • Edit CMakeLists.txt
  • Add set(PRINTING OFF) just above message(STATUS "Printing is ${PRINTING}")
  • Add set(PROFILING OFF) just above message(STATUS "Profiling is ${PROFILING}")
  • Turn DLONG off option (DLONG "Use long integers (64bit) for indexing" OFF)
  • mkdir build
  • cd build
  • cmake .. -DCMAKE_BUILD_TYPE=RELEASE -DCMAKE_INSTALL_PREFIX=~/install -DPYTHON_EXECUTABLE=$(which python3.6) -DPYTHON_STANDARD_LAYOUT=ON (to install in ~/install folder)
  • make install

• Install package that handles the gamepad: pip3 install --user inputs

• Install eiquadprog: sudo apt install robotpkg-eiquadprog

• Install TSID: [https://github.com/stack-of-tasks/tsid#installation] You can put the repo in another folder if you want, like cd ~/install/ instead of cd $DEVEL/openrobots/src/ for the first line.

• Clone interface repository: in /scripts, git clone https://github.com/paLeziart/solopython

Compiling the C++ parts

• Initialize the cmake submodule: git submodule init

• Update the cmake submodule: git submodule udpdate

• Create a build folder: mkdir build

• Get inside and cmake: cd build then cmake .. -DCMAKE_BUILD_TYPE=RELEASE -DCMAKE_INSTALL_PREFIX=~/install -DPYTHON_EXECUTABLE=$(which python3.6) -DPYTHON_STANDARD_LAYOUT=ON

• Compile Python bindings: make

• Copy them to the script folder so that the scripts can access the compiled code: cp python/quadruped_reactive_walking/libquadruped_reactive_walking.so ../scripts/

Run the simulation

• Run python3.6 main_solo12_control.py -i test while being in the scripts folder

• Sometimes the parallel process that runs the MPC does not end properly so it will keep running in the background forever, you can manually end all python processes with pkill -9 python3

Tune the simulation

• In main_solo12_control.py, you can change some of the parameters defined at the beginning of the control_loop function.

• Set envID to 1 to load obstacles and stairs.

• Set use_flat_plane to False to load a ground with lots of small bumps.

• If you have a gamepad you can control the robot with two joysticks by turning predefined_vel to False in main_solo12_control.py. Velocity limits with the joystick are defined in Joystick.py by self.VxScale (maximul lateral velocity), self.VyScale (maximum forward velocity) and self.vYawScale (maximum yaw velocity).

• If predefined_vel = True the robot follows the reference velocity pattern velID. Velocity patterns are defined in Joystick.py, you can modify them or add new ones. Each profile defines forward, lateral and yaw velocities that should be reached at the associated loop iterations (in self.k_switch). There is an automatic interpolation between milestones to have a smooth reference velocity command.

• You can define a new gait in src/Planner.cpp using create_trot or create_walk as models. Create a new function (like create_bounding for a bounding gait) and call it inside the Planner constructor before create_gait_f().

• You can modify the swinging feet apex height in include/quadruped-reactive-control/Planner.hpp with max_height_feet or the lock time before touchdown with t_lock_before_touchdown (to lock the target location on the ground before touchdown).

• For the MPC QP problem you can tune weights of the Q and P matrices in the create_weight_matrices function of src/MPC.cpp.

• Remember that if you modify C++ code you need to recompile the library.