This package contains supplementary code and implementation details for the publication
J. Carius, F. Farshidian and M. Hutter, "MPC-Net: A First Principles Guided Policy Search," in IEEE Robotics and Automation Letters, vol. 5, no. 2, pp. 2897-2904, April 2020.
A preprint is available on arxiv.
While licensing restrictions do not allow us to release the ANYmal model, we are providing our training script with an alternative ball-balancing robot.
-
Build and install Pybind11 according to the instructions in their documentation. Make sure CMake can locate the Pybind11 installation, for example by adding the install path to your
CMAKE_PREFIX_PATH. -
Clone OCS2 into the source folder of a catkin workspace. Then build the python bindings for the optimal control solver with
catkin build ocs2_ballbot_example --cmake-args -DUSE_PYBIND_PYTHON_3=ON -
Install required python packages
pip3 install torch tensorboardX matplotlib
Note that we use python3 as it is required for pytorch.
Make sure your catkin workspace is sourced in the current terminal.
The policy training can then be started with the command
python3 ballbot_learner.py
To monitor progress, execute tensorboard
tensorboard --logdir runs
During training, the policy will be saved to disk in regular intervals.
The performance of the policy on the internal model can be visualized by running the script
python3 ballbot_evaluation.py