ros_pybullet_rl2

An integration between ROS and PyBullet and OpenAI Gym Stable Baselines3 for doing reinforcement learning on a ROS robot in a Pybullet simulation.

Project status

This project is an upgrade of the original ros_pybullet_rl package at (https://github.com/yungchuenng/ros_pybullet_rl_), in line with Stable Baselines 2's upgrade from https://github.com/hill-a/stable-baselines to Stable Baselines 3. It currently presents the following features:

• Omnirobot urdf model (omnirobot_v3).

• The environment and robot scripts for RL training of an Omnirobot for dynamic obstacle avoidance under the Pybullet Gymperium framework.

• ROS sensors plugin into Pybullet robot model: Robot base odometry, laser scanner, camera, force torque sensors.

• ira_laser_tools (citing: Augusto and Pietro at http://wiki.ros.org/ira_laser_tools)

• Adjustable hyperparameters yml files for using the selected RL algorithms as found in src/ros_pybullet_rl2/hyperparams.

• Can be improved:

    - Publishing frequency of the /scan and /d_image topics. 
  

• Limitation:

    - Training can only be done in 1 environment (thread) at a time because of the nature of ROS. 
  

The main overarching training code starts here

Docker

This repository can be installed in its entirety and be ready for use in Docker container by following the instructions in this Docker readme.

• if encounter issue with incompatible CUDA with GPU, see far section below.

Installation

The following instructions have been tested under Ubuntu 16.04 with ROS Kinetic distribution and Ubuntu 18.04 with ROS Melodic distribution.

Clone this repository into your ROS workspace directory, your_workspace_ws/src. Rename this package into "ros_pybullet_rl2".

Use a Python 3.7 environment for the setup.

Install the following dependencies in your setup/environment:

    sudo apt install ros-kinetic-cv-bridge
    sudo apt install ros-kinetic-navigation
    sudo apt install ros-kinetic-eband-local-planner
    sudo apt install ros-kinetic-hector-slam
    sudo apt install make gcc libgtk-3-dev libwebkitgtk-dev libwebkitgtk-3.0-dev freeglut3 freeglut3-dev python-gst-1.0 python3-gst-1.0 libglib2.0-dev ubuntu-restricted-extras libgstreamer-plugins-base1.0-dev
    pip install stable-baselines3[extra]
    pip install -r requirements.txt

Note: If you are using Ubuntu 18.04 and ROS Melodic distribution, the installation for wxpython, cv-bridge, ROS Navigation Stack and Hector SLAM differs:

    sudo apt install ros-melodic-opencv-apps
    sudo apt install ros-melodic-navigation
    sudo apt install ros-melodic-eband-local-planner
    sudo apt install ros-melodic-hector-slam

Check that ros-melodic-cv-bridge is installed and rospkg installed in python3.6 environment. TO VERIFY

To install Pybullet-Gym, clone repository into any desired location and install locally:

    git clone https://github.com/benelot/pybullet-gym.git
    cd pybullet-gym
    pip install -e .

Then do the following:

• Move the directory, ros_pybullet_rl2/common/additions_into_pybullet-gym/yc, into pybullet-gym/pybulletgym/envs.

• Move the file, ros_pybullet_rl2/common/additions_into_pybullet-gym/init.py, into pybullet-gym/pybulletgym/envs.

• Move the directory, ros_pybullet_rl2/common/other_packages/ira_laser_tools, into your_workspace_ws/src.

• Move the directory, ros_pybullet_rl2/common/other_packages/omnirobot_v3, into your_workspace_ws/src.

• Move the directory, ros_pybullet_rl2/common/other_packages/semantic_layers, into your_workspace_ws/src.

• Move the directory, ros_pybullet_rl2/common/additions_into_pybullet-gym/assets, into pybullet-gym/pybulletgym/envs. Copy and merge.

Navigate to your_workspace_ws and build the package:

    catkin_make

To make sure .../your_workspace_ws/src shows in the $ROS_PACKAGE_PATH environment variable, in the same directory, run:

    source devel/setup.bash

To ensure safe loading of configuration file parameters to avoid potential error crashing in between training process, replace .../roslaunch/loader.pywith ros_pybullet_rl2/common/other_packages/roslaunch/loader.py (replace the kinetic with melodic accordingly depending on version) at:

/opt/ros/kinetic/lib/python2.7/dist-packages/roslaunch/loader.py

Run the RL training

• To edit the configuration files for training:

Navigate to ros_pybullet_rl2/config, and edit omnirobot_training_params.yaml (Note NavOmnibase-v1 can only be run with ROS Kinetic). i.e. make sure to change the tensorboard_log input to your desired location to record the log of the reinforcemeant learning training process.

Run the training:

    roslaunch ros_pybullet_rl2 nav_train.launch

• To monitor the training of your robot at any instance in time (Note this is not real-time), navigate to src/ros_pybullet_rl2 and run:

    python check_training_progress.py --model-path logs/ppo2/Omnibase-v1_1
  

Change the model path according to the path to your model directory or log path. The trained model is saved in src/ros_pybullet_rl2/logs/algorithm_name/env_id by default.

• To visualise the robot agent in rviz:

    rosrun rviz rviz
  

This is possible, but it may cause the computer to freeze if there is insufficient memory space or the CPU usage becomes fully clocked. This is able to visualise the robot model and the /scan, /rgb_image, /odom topics. Select the relevant topics for visualisation.

• To load a previously trained agent for further training:

Navigate to ros_pybullet_rl2/config, and edit omnirobot_training_params.yaml. Change the trained_agent parameter to the relative path (from ros_pybullet_rl2.py) of your trained agent directory, i.e. logs/ppo2/Omnibase-v1_1.

• The default hyperparameters used for training may also be override by specifying in the omnirobot_training_params.yaml. To change the neural network architecture i.e. from 2 layers of 64 nodes to 4 layers of 64 nodes, edit net_arch=[dict(pi=[64, 64], vf=[64, 64])] to net_arch=[dict(pi=[64, 64, 64, 64], vf=[64, 64, 64, 64])]:

hyperparams:                {'policy': 'MlpPolicy',
                              'policy_kwargs': "dict(log_std_init=-2,
                               ortho_init=False,
                               activation_fn=nn.ReLU,
                               net_arch=[dict(pi=[64, 64], vf=[64, 64])]
                               )"
}

• To change the input data mode to the neural network, edit the obs_input_type parameter (take note that the RL environment has to support the use of this parameter). The choices are ['default','array', 'multi_input']. 'array' ('defaullt') is for use with MlpPolicy, 'multi_input' is for use with MultiInputPolicy.

obs_input_type:             array

• To choose a simulation environment for training:

Navigate to config/omnirobot_training_params.yaml. Edit the select_env parameter according to the number ID of the preset environment you would like to train the robot in.

0. Stadium maze environment
1. Simple straight line
2. Curvy corridor
3. Twisting corridor
4. Straight line with dynamic obstacles
5. T-junction with unseen static obstacles
6. T-junction with dynamic obstacles
7. Hall, goals in circular formation
8. Static environment with deadends
9. Dynamic obstacles environment without static obstacles
10. Lab
11. Lab (Dynamic)
12. Doorway

As of now, the average rate at which the relevant sensors and output command are published for use in reinforcement learning are recorded as follows:

    - /depth_image_processed: 0.420 Hz 
    - /contact_sensor: 42 Hz 
    - /scan: 5.3 Hz 
    - /odom: 42 Hz 
    - /cmd_vel: 55 Hz

To set the number of timesteps at which to do an evaluation of performance (not policy evaluation) (i.e. reward mean) of the agent in training:
- n_steps in ros_pybullet_rl2/src/ros_pybullet_rl2/hyperparams/{algorithm_name}.yml

To set the number of timesteps at which to do a policy evaluation for the training agent: - eval_freq in config/omnirobot_training_params.yaml

• To fix the set of goals to be achieved by the robot agent in each environment:

Navigate to config/env_pybullet_params.yaml and under each environment, the goal_set variable consisting of a list of tuples may be edited to change the set of goals to be reached for the learning agent. If goal_set is set to '[]', random goals will be generated in the environment.

Run the validation of trained models

First, place the trained model, i.e. Omnibase-v1_1, found in src/ros_pybullet_rl2/logs/algorithm_name directory (by default), into src/ros_pybullet_rl2/trained_agents/allgorithm_name directory, for example:

sudo cp -r /root/workspace_ws/src/ros_pybullet_rl2/src/ros_pybullet_rl2/logs/ppo /root/workspace_ws/src/ros_pybullet_rl2/src/ros_pybullet_rl2/trained_agent

• To edit the configuration files for validation:

Navigate to ros_pybullet_rl2/config, and edit omnirobot_validation_params.yaml. Note that the load_best parameter when set to True indicates that the best_model.zip in the trained model directory will be loaded for validation, else the latest trained model.zip will be loaded instead. Set the number of timesteps to run the validation program by changing the n_timesteps parameter as required.

Run the validation program:

    roslaunch ros_pybullet_rl2 nav_validation.launch

• To choose an environment for validation:

Navigate to config/omnirobot_validation_params.yaml. Edit select_env parameter according to the number ID of the preset environment you would like to train the robot in.

0. Stadium maze environment
1. Simple straight line
2. Curvy corridor
3. Twisting corridor
4. Straight line with dynamic obstacles
5. T-junction with unseen static obstacles
6. T-junction with dynamic obstacles
7. Hall, goals in circular formation
8. Static environment with deadends
9. Dynamic obstacles environment without static obstacles
10. Lab
11. Lab (Dynamic)
12. Doorway

• To fix the set of goals to be achieved by the robot agent in each environment:

Navigate to config/env_pybullet_params.yaml and under each environment, the validation_goal_set variable consisting of a list of tuples may be edited to change the set of goals to be reached for the trained agent. If goal_set is set to '[]', random goals will be generated in the environment.

Visualise training process

• To visualise the training process using Tensorboard visualisation toolkit by TensorFlow:

According to the tensorboard_log directory location, i.e. /home/ubuntu/your_directory_name, as set in "Test the RL training", navigate to the directory containing this directory and run:

    tensorboard --logdir your_directory_name

Following the instructions as printed on the terminal, run the given link on a browser to start up the Tensorboard window, for instance, http://my_computer_name:6006.

Pybullet robot agent sensors retrofitting

• To choose the relevant sensors for initialisation on the simulation robot model:

Navigate to config/omnirobot_pybullet_params.yaml. Here, the plugin sensors to fit onto the Pybullet robot model may be introduced into the plugins parameter. The relevant sensor information will have to set accordingly below, for LiDAR, RGBD Camera, force-torque sensors etc. The rate at which the Pybullet simulation environment runs may also be controlled via the parameter loop_rate.

• To introduce a new sensor plugin to the robot model:

Navigate to src/ros_pybullet_rl2/plugins. The sensor plugins equipped onto the robot is found here. Refer to plugin_template.py for a template to create a custom plugin.

Standard Pybullet environment construction

The Pybullet simulation environment is created based on the createMultiBody function. The parameters of the generated bodies are written in config/env_pybullet_params.yaml.

• To create new environments for simulation:

Navigate to env_pybullet_params.yaml, imitating the format used to construct the existing standard environments, existing environments may be edited and new environments may be created.

Incompatible Pytorch version with device GPU CUDA error

According to your computer GPU device, do the following steps (or refer to this website, but according to your own device GPU (this example uses NVIDIA GeForce RTX 3070 GPU):

On a python terminal, it originally showed:

torch.cuda.get_arch_list()
[‘sm_37’, ‘sm_50’, ‘sm_60’, ‘sm_70’]
torch.version.cuda
‘10.2’

But sm_86 is required for RTX3070, therefore, will require a version of Pytorch with i.e. cuda 11.4 wheel, doing the following installation should solve the issue:

pip3 install torch torchvision torchaudio --extra-index-url https://download.pytorch.org/whl/cu114

Now, the python terminal, would show:

torch.cuda.get_arch_list()
['sm_37', 'sm_50', 'sm_60', 'sm_70', 'sm_75', 'sm_80', 'sm_86']
torch.version.cuda
'11.7'

Enjoy!