Explore-Bench

Explore-Bench is developed aiming to evaluate traditional frontier-based and deep-reinforcement-learning-based autonomous exploration approaches in a unified and comprehensive way.

The related paper "Explore-Bench: Data Sets, Metrics and Evaluations for Frontier-based and Deep-reinforcement-learning-based Autonomous Exploration" is accepted to the 2022 International Conference on Robotics and Automation (ICRA 2022).

Features:

Data Sets: various basic exploration scenarios (i.e., loop, narrow corridor, corner, and multiple rooms) and their combinations are designed.
Metrics: two types of metrics (efficiency metrics and collaboration metrics) are proposed.
Platform: a 3-level platform with a unified data flow and $12 \times$ speed-up that includes a grid-based simulator for fast evaluation and efficient training, a realistic Gazebo simulator, and a remotely accessible robot testbed for high-accuracy tests in physical environments is built.
Evaluations: one DRL-based and three frontier-based exploration approaches are evaluated and some insights about the selection and design of exploration methods are provided.

Dependency

The project has been tested on Ubuntu 18.04 (ROS Melodic). To run the exploration approach on Desktop PC or real robots, please install these packages first:

Cartographer for map building

Cartographer is a 2D/3D map-building method. It provides the submaps' and the trajectories' information when building the map. We slightly modified the original Cartographer to make it applicable to multi-robot SLAM and exploration. Please refer to Cartographer-for-SMMR to install the modified Cartographer to carto_catkin_ws and

source /PATH/TO/CARTO_CATKIN_WS/devel_isolated/setup.bash

MAPPO for reinforcement learning training and evaluation

MAPPO is a multi-agent variant of PPO (Proximal Policy Optimization), which is a SOTA on-policy reinforcement learning algorithm. We slightly modified the original marlbenchmark/on-policy and provide the training and test code in this repo.

# create conda environment
conda create -n marl python==3.8.10
conda activate marl
pip install torch torchvision

# install onpolicy package
cd onpolicy
pip install -e .

We provide requirement.txt but it may have redundancy. We recommend that the user try to install other required packages by running the code and finding which required package hasn't installed yet.

Turtlebot3 Description and Simulation

sudo apt install ros-melodic-turtlebot3*
sudo apt install ros-melodic-bfl
pip install future
sudo apt install ros-melodic-teb-local-planner
echo 'export TURTLEBOT3_MODEL=burger' > ~/.bashrc

After installing these dependencies, put these packages in your ROS workspace (i.e. catkin_ws/src) and catkin_make.

Other Dependencies

Except for the DRL-related scripts (running under Python3.8), we use C++ and Python2.7 to implement the benchmark.

There are some packages to be installed:

pip install autolab_core==0.0.14

Run Frontier-based Exploration in Gazebo (Level-1)

Template:

{env}      = 'loop' or 'corridor' or 'corner' or 'room' or 'loop_with_corridor' or 'room_with_corner'
# for multiple robots, there are two env cases: 'env_close' and 'env_far' (i.e. 'loop_close' and 'loop_far')
{number_robots} = 'single' or 'two'
{method}        = 'rrt' or 'mmpf' or 'cost'
{suffix}        = 'robot' or 'robots' (be 'robot' when number_robots != 'single')
# build simulation environment
roslaunch sim_env {env}_env_{number_robots}_{suffix}.launch
# start cartographer map building and move_base
roslaunch sim_env {number_robots}_{suffix}.launch
# start frontier-based exploration
roslaunch exploration_benchmark {number_robots}_{method}_node.launch
# start data logging and evaluating the exploration performance
cd exploration_benchmark/scripts
python exploration_metric_for_{number_robots}_{suffix}.py '../blueprints/{env}.pgm' '../blueprints/{env}.yaml'

For example,

Room Environment -- Single Robot -- Field-based Exploration (MMPF)

roslaunch sim_env room_env_single_robot.launch
roslaunch sim_env single_robot.launch
roslaunch exploration_benchmark single_mmpf_node.launch

Then, start a new terminal and use our proposed metrics to evaluate the exploration performance:

cd exploration_benchmark/scripts
python exploration_metric_for_single_robot.py '../blueprints/room.pgm' '../blueprints/room.yaml'

At last, choose "Publish Point" button in the rviz and then click anywhere in the map to start the exploration.

Corridor Environment (Far) -- Two Robots -- Cost-based Exploration

roslaunch sim_env corridor_far_env_two_robots.launch
roslaunch sim_env two_robots.launch
roslaunch exploration_benchmark two_cost_node.launch

Then, start a new terminal and use our proposed metrics to evaluate the exploration performance:

cd exploration_benchmark/scripts
python exploration_metric_for_two_robots.py '../blueprints/corridor.pgm' '../blueprints/corridor.yaml'

Run DRL-based Exploration in Gazebo (Level-1)

To run the DRL-based exploration, we need Python3 and conda environment.

Single Robot

First, start the simulation environment , mapping module, rviz visualization and performance evaluation,

roslaunch sim_env room_env_single_robot.launch
roslaunch sim_env single_robot.launch
roslaunch exploration_benchmark single_rl_node.launch
cd exploration_benchmark/scripts
python exploration_metric_for_single_robot.py '../blueprints/room.pgm' '../blueprints/room.yaml'

Then, start a new terminal and run the DRL-based exploration,

# enter conda env
source ~/anaconda3/bin/env marl
cd exploration_benchmark/scripts/DRL
# run the DRL model
bash run_single_robot.sh

Multiple Robots

First, start the simulation environment , mapping module, rviz visualization and performance evaluation,

roslaunch sim_env room_far_env_two_robots.launch
roslaunch sim_env two_robots.launch
roslaunch exploration_benchmark two_rl_node.launch
cd exploration_benchmark/scripts
python exploration_metric_for_two_robots.py '../blueprints/room.pgm' '../blueprints/room.yaml'

Then, start a new terminal and run the DRL-based exploration,

# enter conda env
source ~/anaconda3/bin/env marl
cd exploration_benchmark/scripts/DRL
# run the DRL model
bash run_two_robots.sh

Note: you need to choose "Publish Point" button in the rviz and then click anywhere in the map to start the performance evaluation.

Train DRL Model in Grid-based Simulator (Level-0)

Explore-Bench supports the DRL training in a fast grid-based simulator (Level-0).

Follow these steps to train your own DRL model:

# enter conda env
source ~/anaconda3/bin/env marl
# ensure that you have installed the MAPPO dependency
cd onpolicy/onpolicy/scripts
# run the DRL training 
bash train_grid.sh

The training parameters can be modified according to the user's need, i.e., the number of robots (num_agents), the hidden size, the batch size and so on.

Refer to train_grid.sh for details.

Evaluate Exploration Approaches in Grid-based Simulator (Level-0)

Besides training DRL models, the grid-based simulator can be used for fast evaluation of both frontier-based and DRL-based methods.

The field-based and cost-based exploration approaches are taken for example:

# enter conda env
source ~/anaconda3/bin/env marl
cd grid_simulator
# evaluate cost in corner env (2 robots)
python GridEnv.py cost 2 ../onpolicy/onpolicy/envs/GridEnv/datasets/corner.pgm
# evaluate cost in corner env (1 robots)
python GridEnv.py cost 1 ../onpolicy/onpolicy/envs/GridEnv/datasets/corner.pgm
# evaluate mmpf in corner env (2 robots)
python GridEnv.py mmpf 2 ../onpolicy/onpolicy/envs/GridEnv/datasets/corner.pgm
# evaluate mmpf in corner env (1 robots)
python GridEnv.py mmpf 1 ../onpolicy/onpolicy/envs/GridEnv/datasets/corner.pgm

Citation

If you find our work useful in your research, please consider citing:

@article{xu2022explore,
  title={Explore-Bench: Data Sets, Metrics and Evaluations for Frontier-based and Deep-reinforcement-learning-based Autonomous Exploration},
  author={Xu, Yuanfan and Yu, Jincheng and Tang, Jiahao and Qiu, Jiantao and Wang, Jian and Shen, Yuan and Wang, Yu and Yang, Huazhong},
  journal={arXiv preprint arXiv:2202.11931},
  year={2022}
}

YongzhengCui/Explore-Bench