/safe_rl_manipulators

Verifiably Safe Deep Reinforcement Learning for Robotic Manipulationin Human Environments

Primary LanguageC++MIT LicenseMIT

Robot RL

Deep reinforcement learning (RL) has shown promising results in robotic manipulator path planning for complex goals. However, to this point, no method guarantees the safety of highly dynamic obstacles, such as humans, in manipulator control. This lack of formal safety prevents the application of RL for manipulators in real-world human environments. Therefore, we propose a shielding mechanism that guarantees ISO-verified human safety while training and deploying RL algorithms on manipulators. We utilize a fast reachability analysis of humans and manipulators to guarantee that the manipulator comes to a full stop before a human can reach it. Our proposed method not only guarantees safety but also significantly improves the RL performance by preventing episode-ending collisions. We present the performance of our proposed method in simulation using human motion capture data.

Human reachability analysis

Our code is partly based on an early version of our published human reachability package SaRA. We are planning on integrating the new version of SaRA in our safety shield soon.

Installation

We work on Ubuntu 20.04. If you run Ubuntu 20.04, you can install our project locally with the following steps. Otherwise, you can use a Docker container as described below. A Nvidia graphics card is required if you want to display the simulation GUI.

  1. Install ROS noetic http://wiki.ros.org/noetic/Installation/Ubuntu

  2. Go into catkin workspace root folder cd src/catkin_ws

  3. Copy your PATH_TO_CATKIN_WS:

    pwd

  4. Add to your ~/.bashrc:

    source /opt/ros/noetic/setup.bash
export ROBOT_RL_SIM_ROOT=PATH_TO_CATKIN_WS
alias kill_gazebo='$ROBOT_RL_SIM_ROOT/kill_all_processes.sh'

    Make sure to replace PATH_TO_CATKIN_WS with the copied path. The kill_gazebo command can be used if not all ros and gazebo tasks are closed.

  5. Install anaconda https://docs.anaconda.com/anaconda/install/linux/) and create a conda environment with Py 3.6

    conda create -n robot-rl python=3.6
conda activate robot-rl

  6. Go into catkin workspace root folder cd src/catkin_ws

  7. Install ros dependencies. Do this now and when checking out a new branch or pulling big changes:

    rosdep update
rosdep install --from-paths src --ignore-src -r -y

  8. Install pip requirenments

    pip3 install -r requirements.txt --upgrade

  9. (Optional): Install spinup.

    cd src/spinningup/
sudo apt-get update && sudo apt-get install libopenmpi-dev
pip install -e .

    Unfortunately, the spinningup repo has a bug regarding tf1 and pytorch. There is already a ticket on their github and I commented a simple solution (openai/spinningup#333).

  10. Compile the modrob workstation library To compile the workstation library, execute the following steps:

    1. create a build directory in the workstation folder and cd into it
    mkdir $ROBOT_RL_SIM_ROOT/src/modrob_workstation/workstation/build
cd $ROBOT_RL_SIM_ROOT/src/modrob_workstation/workstation/build
    1. generate a make file with cmake
    cmake ..
    1. compile the library with the generated make file
    make

  11. Build project In catkin workspace root folder: Either

    ./debug_build.sh

    for debug and development, or

    ./clean_build.sh

    for deployment and training on server. IMPORTANT:

    source devel/setup.bash

    This needs to be repeated for every new bash.

Modular robot simulation

Run the training

roslaunch example_modrob_human_training start_training_safe_human_her_sac_modrob1.launch

Important configuration files

Set the parameters of the RL algorithm in

src/catkin_ws/src/example_modrob_human_training/config/my_modrob_safe_human_params_sac_her_modrob1.yaml

Set the parameters of the robot task (like goal position, accuracy, time in between movements, etc.) in

src/catkin_ws/src/openai_ros/openai_ros/src/openai_ros/task_envs/human_modrob/config/modrob_safe_human_random_modrob1.yaml

You can use the recommended parameters for the random goal experiment in

src/catkin_ws/src/openai_ros/openai_ros/src/openai_ros/task_envs/human_modrob/config/modrob_safe_human_random_modrob1_all_rand.yaml

or for the evading human experiment in

src/catkin_ws/src/openai_ros/openai_ros/src/openai_ros/task_envs/human_modrob/config/modrob_safe_human_random_modrob1_single_goal.yaml

To disable the safety shield, go to the file

src/catkin_ws/src/openai_ros/openai_ros/src/openai_ros/robot_envs/modrob_env_path_following.py

and change the line 110 from

launch_file_name="init_modrob.launch",

to

launch_file_name="init_modrob_unsafe.launch",

The results will be saved in

src/catkin_ws/src/example_modrob_human_training/training_results

Test pretrained models

We provide the training results of the safe and unsafe agent (baseline) for two experiments: random goal and evading human. These experiments are located in

src/catkin_ws/src/example_modrob_human_training/training_results

To test these models, you will need to adapt the file

src/catkin_ws/src/example_modrob_human_training/config/my_modrob_safe_human_params_sac_her_modrob1_load.yaml

Choose the correct outdir and choose the load_epoch (195 for fully trained model).

Docker setup

Without GUI (for training on server)

Install docker

Install docker according to https://docs.docker.com/engine/install/ubuntu/

Build the docker image

Build the docker container (You only need to rebuild if you change something in the package.xml files or in the spinningup folder.)

cd src/catkin_ws
./build_docker.sh

Run the docker image without graphical output

./run_docker_no_gui.sh

Build the ros project (see below)

With GUI (for local machine)

Since the simulation needs the graphical representation of gazebo, we need to run a nvidia-docker2 container. A nvidia graphics card with CUDA support is required.

Install nvidia cuda toolkit

sudo apt-get install nvidia-cuda-toolkit This is not the deep learning support. Check cuDNN for this.

Install nvidia docker

curl -s -L https://nvidia.github.io/nvidia-docker/gpgkey | sudo apt-key add -
distribution=$(. /etc/os-release;echo $ID$VERSION_ID)
curl -s -L https://nvidia.github.io/nvidia-docker/$distribution/nvidia-docker.list | \
  sudo tee /etc/apt/sources.list.d/nvidia-docker.list
sudo apt-get update
sudo apt-get install -y nvidia-docker2
sudo pkill -SIGHUP dockerd

Build the docker image

Build the docker container (You only need to rebuild if you change something in the package.xml files or in the spinningup folder.)

cd src/catkin_ws
./build_docker.sh

Run the docker image with graphic output

./run_docker_gui.sh

Build the ROS project

Execute every command in ./docker_build_ws.sh manually:

export PYTHONPATH="${PYTHONPATH}:$ROBOT_RL_SIM_ROOT/src/spinningup/"
sudo rm -r devel/ build/
conda activate robot-rl
pip3 install -r requirements.txt --upgrade
catkin_make clean
mkdir $ROBOT_RL_SIM_ROOT/src/modrob_workstation/workstation/build
cd $ROBOT_RL_SIM_ROOT/src/modrob_workstation/workstation/build
cmake ..
make
cd $ROBOT_RL_SIM_ROOT
catkin_make
source devel/setup.bash
rospack profile