opt-mimic-robot-deploy

Robot deployment code for the paper “OPT-Mimic: Imitation of Optimized Trajectories for Dynamic Quadruped Behaviors”. This repo was based on the ODRI Solo repo, rewritten to exclude unused features (eg. integration with dynamic_graph_manager and TI hardware), and to support custom functionality (eg. control through neural networks).

Setup Instructions

Note: these instructions are not maintained for compatibility with newer versions of external software (eg. NYU/MPI packages).

Enviroment setup

First, an installation of Ubuntu 18 with the RT_PREEMPT kernel patch must be set up, with ROS2 also installed. The exact steps that I performed are documented on this google doc (last tested Fall 2021).

Install dependencies and set up project directory

1. mkdir and cd into a directory to contain all project files. In my case ~/Documents/robot_script_ws
2. Clone the Machines in Motion Lab treep package. In my case

git clone git@github.com:machines-in-motion/treep_machines_in_motion.git

1. Use treep to clone the real_time_tools repo and all of its dependencies:

treep --clone REAL_TIME_TOOLS

Do the same for the odri_control_interface repo:

treep --clone ODRI_CONTROL_INTERFACE

Afterwards, my directory looks like

robot_script_ws/
	treep_machines_in_motion/
	workspace/
		src/
			googletest/
			master-board/
			mpi_cmake_modules/
			odri_control_interface/
			pybind11/
			real_time_tools/
			yaml_utils

1. Download LibTorch from the PyTorch website. The options I chose were:

Additionally, I chose the cxx11 ABI. Note that since this machine is running an rt-preempt kernel, Nvidia drivers are unsupported for it and the device should be CPU (and not CUDA).

1. Extract the zip file and move the resulting libtorch folder to the current directory. Then my directory becomes

robot_script_ws/
	libtorch/
		bin/
		build-hash
		build-version
		include/
		lib/
		share/
	treep_machines_in_motion/
	workspace/
		src/
			googletest/
			master-board/
			mpi_cmake_modules/
			odri_control_interface/
			pybind11/
			real_time_tools/
			yaml_utils

1. Create directories config, models, and traj. These respectively store joint calibration files, neural network models, and reference trajectory csv files.

mkdir config models traj

Now my directory is

robot_script_ws/
	config/
	libtorch/
		bin/
		build-hash
		build-version
		include/
		lib/
		share/
	models/
	traj/
	treep_machines_in_motion/
	workspace/
		src/
			googletest/
			master-board/
			mpi_cmake_modules/
			odri_control_interface/
			pybind11/
			real_time_tools/
			yaml_utils

1. Clone this repo into the colcon src directory

cd workspace/src/
git clone https://github.com/yunifuchioka/opt-mimic-robot-deploy.git

Finally, my directory becomes

robot_script_ws/
	config/
	libtorch/
		bin/
		build-hash
		build-version
		include/
		lib/
		share/
	models/
	traj/
	treep_machines_in_motion/
	workspace/
		src/
			googletest/
			master-board/
			mpi_cmake_modules/
			odri_control_interface/
			opt-mimic-robot-deploy/
			pybind11/
			real_time_tools/
			yaml_utils

Build and Source the Project

These steps have to be run for every new terminal, and steps 3 and 4 have to be run after any modifications to C++ code.

1. Switch to root, which is necessary for the network communcation

sudo -s

1. Source ROS2. In my case

source /opt/ros/dashing/setup.bash
source /opt/openrobots/setup.bash

1. cd to the workspace directory (cd ~/Documents/robot_script_ws/workspace in my case), then build with colcon. I belive that the arguments to specify the libtorch installation location only needs to be run once on initial project compilation (ie just colcon build for later builds). If the build fails, it’s most likely because of changes to cmake files in external packages that haven’t been accounted for. In this case, I look for the chunk of cmake code corresponding to the error message, and comment out the problematic lines (a bit hacky, but it’s likely not worth debugging cmake code accross multiple packages…).

colcon build --cmake-args -DCMAKE_PREFIX_PATH=$PWD/../libtorch

1. Source the setup file

source install/setup.bash

Import Neural Network Policies

In order for a pytorch neural network model produced by RL training to be readable from a C++ program, it must first be converted to a torchscript model. See the Pytorch tutorial on loading a torchscript model in C++ for more details. Here are the steps to perform this conversion:

1. Copy the trained model file, say my_trained_model.pt for example, into the models folder. The directory should look like

robot_script_ws/
	config/
	libtorch/
	models/
		my_trained_model.pt
	traj/
	treep_machines_in_motion/
	workspace/

1. From a non-sudo user, run

	python src/opt-mimic-robot-deploy/srcpy/convert_torchscript_model.py my_trained_model
	

I recommend doing this within a python virtual environment with pytorch installed. Also note that the string argument is the name of the model file, without the `.pt` file extension.

Afterwards, my directory looks like

	robot_script_ws/
		config/
		libtorch/
		models/
			my_trained_model.pt
			my_trained_model_script.pt
		traj/
		treep_machines_in_motion/
		workspace/
	

where `my_trained_model_script.pt` is the equivalent model, converted to a torchscript model so it can be loaded from C++ 3. Then the RL policy to be used is specified in `main.cpp` with the command `controller.initialize_network("my_trained_model");`.

Import Reference Trajectory Files

Since the trained RL policy is trained to output residual position targets, the reference motion trajectory file must also be used to execute the motion on the robot.

1. Copy the reference csv file produced from trajectory optimization onto the traj folder.
2. Then the reference motion to be used is specified in main.cpp with the command ref_traj = openData("../traj/my_ref_traj.csv");.

Calibrate Joint Angles

1. Run

ros2 run robot_script calibrate MY_INTERFACE

where MY_INTERFACE if the name of the interface, obtained from running ifconfig.

Run Main Script

1. Run the script

ros2 run robot_script main MY_INTERFACE

where MY_INTERFACE if the name of the interface, obtained from running ifconfig. Within main.cpp, the lines ref_traj = openData("../traj/my_ref_traj.csv"); and controller.initialize_network("my_trained_model"); must have arguments corresponding to the reference trajectory file and neural network file being used.