The package provides a framework to generate 3D velocity commands directly from a depth image stream. A part of this work is currently under review and has two modes 'apf' and 'qmdp'. The discussion about the 'apf' mode is provided in the paper

@inproceedings{ahmad2021apf,
  title={Apf-pf: Probabilistic depth perception for 3d reactive obstacle avoidance},
  author={Ahmad, Shakeeb and Sunberg, Zachary N and Humbert, J Sean},
  booktitle={2021 American Control Conference (ACC)},
  pages={32--39},
  year={2021},
  organization={IEEE}
}

'qmdp' mode is detailed in the article

@article{ahmad2021end,
  title={End-to-End Probabilistic Depth Perception and 3D Obstacle Avoidance using POMDP},
  author={Ahmad, Shakeeb and Sunberg, Zachary N and Humbert, J Sean},
  journal={Journal of Intelligent \& Robotic Systems},
  volume={103},
  number={2},
  pages={1--18},
  year={2021},
  publisher={Springer}
}

~pt_cloud_in: Ordered pointcloud representation of depth images in a depth camera's optical frame
~cam_info_in: Camera info topic from the depth camera

~viz_out: Vizualization topic
~pt_out: Most probable location of the occupied region inside the depth image stream
~twist_out: Repulsive velocity out (in qmdp mode)
~force_out: Repulsive force out (in apf mode)

~read_from_file: Read transition, reward and alpha vector matrices from a file
~matrices_file_path: Folder to save the transition, reward and alpha vector matrices
~n_threads: Number of threads to run multi-thread (OpenMP is required to to use multiple threads)
~distance_interval: Discretization interval
~pixel_interval: Discretization interval
~min_distance: Mininimum depth horizon
~max_distance: Maximum depth horizon
~n_particles_vox: Number of particles
~trans_noise_sdev: [\sigma_s, \sigma_z]
~obsv_noise_sdev: [\sigma_o, \sigma_n]

~max_action_values: Maximum velocity in x,y and z directions (only in qmdp mode)
~min_action_values: Minimum velocities in x,y and z directions (only in qmdp mode)
~action_intervals: Discretization intervals for the velocities in x,y and z directions (only in qmdp mode)

~reward_Q: Reward Q matrix (diagonal) (only in qmdp mode)
~repulsive_potential_max_distance: Maximum effective repulsive potential distance
~repulsive_potential_gain: Repulsive potential gain

~alpha_vector_iterations: Number of iterations for alpha vector update (only in qmdp mode)

~lookahead_time: Lookahead time to calculate the reward (only in qmdp mode)
~sampling_time: Sampling time for the lookahead path to calculate the reward (only in qmdp mode)
~base_frame_id: Robot body frame id

~output_apf: True enables apf mode
~output_qmdp: True enables qmdp mode

~publish_viz: True enables RViz visualization msg out
~display_cout: True enables verbose

~pose_in: Robot position in.
~goal_pt_in: Goal point in.
~rep_vec_in: Repulsive velocity/force in from filter_point_node.

~twist_out: Final velocity action for the robot.

~att_vel: Forward vertical and yaw rates for the attractor pull
~rep_vec_timeout_secs: Timeout for repulsive vector
~process_rate_secs: Final action publish rate
~success_radius: Success radius
~holonomic: False enables non-holonomic
~yaw_err_bound_nonzero_fore_vel_in_rad: No forward velocity if the yaw error is greater than this value. Set to 1.57 to avoid negative forward velocities

~mode: 'apf' or 'qmdp'
~parabolic_attractor_bound: Radius around the goal to enable parabolic attractor (only in apf mode)
~attractor_gain: Attractor gain (only in apf mode)

~display_cout: True enables verbose

~pose_frame_id: Frame id of the incoming pose

~base_frame_id: Robot body frame id
~pose_frame_output: True enables the output in pose/world frame

3.1 Install ROS with RViz and Gazebo by following the instructions on this link. The package is tested on ROS Melodic and Ubuntu 18.04.5 LTS.
3.2 Create a ROS workspace as follows

$ mkdir -p ~/catkin_ws/src
$ cd ~/catkin_ws/src

3.3 Install rotors_simulator package from https://github.com/ethz-asl/rotors_simulator
3.4 Install depth_image_proc package from apt sudo apt install ros-melodic-depth-image-proc or from source (http://wiki.ros.org/depth_image_proc)

3.5 Clone the package and its dependencies

$ cd ~/catkin_ws/src
$ git clone -b apf_pf_doc https://github.com/shakeebbb/apf_pf
$ git clone -b apf_pf_doc https://github.com/shakeebbb/depth_noiser
$ git clone -b apf_pf_doc https://github.com/shakeebbb/ros_conversions
$ git clone -b apf_pf_doc https://github.com/shakeebbb/drone_pose

3.6 Build the workspace using python_catkin_tools (you may need to install python_catkin_tools)

$ cd ~/catkin_ws/
$ catkin build

3.7 Source the ROS workspace

$ echo "source ~/catkin_ws/devel/setup.bash" >> ~/.bashrc
$ source ~/.bashrc

4.1 Open a terminal and run the simulator

$ roslaunch apf_pf rotors_sim.launch

4.2 Open a new terminal and run apf_pf

$ export ROS_NAMESPACE=iris
$ export TOF_FORE_NAME=vi_sensor
$ roslaunch apf_pf rotors_apf_pf.launch launch_depth_noiser:=false

To inject noise into the depth image pixels, launch apf_pf above with launch_depth_noiser:=true and set constant_noise, linear_noise and quadratic_noise parameters inside the rotors_apf_pf.launch to the desired values.
Note: Depth noiser adds latency to the depth information depending on the hardware it is being run on.
4.3 Open a new terminal and run velocity controller for the rotors_simulator

$ export ROS_NAMESPACE=iris
$ roslaunch drone_pose rotors_vel_control.launch

4.4 Open a new terminal and run rviz to checkout the visualizations

$ rosrun rviz rviz -d ~/catkin_ws/src/apf_pf/rviz/rotors_config.rviz

4.5 Open a new terminal and publish a desired goal point for the robot. IMPORTANT: Wait for the next message to appear after [ INFO] [xxxx.xx, xxxx.xxx]: Computing alpha vectors ... inside the terminal running apf_pf, before sending the goal point.

$ rostopic pub -r 10 /iris/lookahead_point geometry_msgs/PointStamped "header:
  seq: 0
  stamp:
    secs: 0
    nsecs: 0
  frame_id: 'world'
point:
  x: 100.0
  y: 0.0
  z: 1.5"

4.6 An instance of the simulation run obtained following the preceding steps is shown here for reference.