quadruped-rl

Implementation of a data-based control architecture on the Solo-12 quadruped, using neural networks trained with reinforcement learning.

Training

See [https://github.com/Gepetto/soloRL] for more details about the training environment.

Dependencies

• Install a Python 3 version (for instance Python 3.8)

• Install PyBullet (simulation environment): pip3 install --user pybullet

• Install Inputs (to control the robot with a gamepad): pip3 install --user inputs

• Clone git clone --recursive https://gitlab.laas.fr/paleziart/quadruped-rl.git

• Compile the C++ part that runs the neural network:

  • cd cpuMLP
  • mkdir build
  • cd build
  • cmake .. -DCMAKE_BUILD_TYPE=RELEASE -DPYTHON_EXECUTABLE=$(which python3) -DPYTHON_STANDARD_LAYOUT=ON
  • make

• Run the controller in simulation: python3 main_solo12_RL.py

• You can make the simulation longer by tweaking the max_steps of RLParams in Params.py. The control loop runs at 1 kHz so the number is in milliseconds.

• You can disable the plots at the end of the simulation by setting to False the variable PLOTTING of RLParams in Params.py.

Using a joystick

• You can control the robot with your gamepad turning USE_PREDEFINED to False and USE_JOYSTICK to True.

• You will likely have to tweak the following formulas in Joystick.py since the values returned for each axes depends on the model of your gamepad.

self.vX = (self.gp.leftJoystickX.value / 0.00390625 - 1 ) * self.VxScale
self.vY = (self.gp.leftJoystickY.value / 0.00390625 - 1 ) * self.VyScale
self.vYaw = (self.gp.rightJoystickX.value / 0.00390625 - 1 ) * self.vYawScale

Deployment on a real Solo-12

• TODO