This node controls a robot to look for a gas source, using the infotaxis algorithm. For detail about infotaxis itself, see neurovertex/infotaxis
This node needs a map and a location. The two obvious approaches are map_server + AMCL or SLAM. This node will also need the navigation stack (move_base), and won't run without a global costmap and a local costmap being published.
If not in simulation (which is so far the only implemented mode, actual sensors reading coming soon), you'll need a gas sensor mounted on a mobile robot. That's about it. Format of the sensor readings is TBD yet.
The infotaxis algorithm is obviously not implementable "as is" on a real robot. Special measures have to be taken to account for obstacles, the continuous nature of the robot's position, moving times, etc... So here are the main differences between the algorithm and this implementation.
The node is implemented as a finite-state automaton. States are DETECTING, PLANNING, MOVING and EXIT. Entry state is DETECTING and exit state is ... obvious enough.
When reaching this state, the robot will wait for dt seconds while reading from the sensor as to get a number of detections (as per the algorithm). How this is done with a real sensor is yet TBD. In Simulation mode, the detections will be randomly generated according to a Poisson distribution. Since the robot spends much more time moving between goals than arrived, most of the detections are done during the moving state. Then the sate is set to PLANNING.
Here the nodes determines where to go next. It generates a number of poses around its current location, then picks the one with the highest expected entropy reduction, and sends it to the navigation stack. Since calculating said values takes some time, most of the planning is actually done during the moving state.
Every step in the MOVING state, nextsteps new goals are generated. They are spread around the current destination by steps of 360/nextsteps° (i.e 90° for nextsteps=4). The distance is random, and its min and max possible values vary depending on on the "interest" of the area the destination is in, the interest is calculated by taking the N~th root of the average value of the grid around the center (N ~ 25). The actual boundaries for the distance are then calculated from the parameters min_stepdist, med_stepdist, max_stepdist; the lower and upper bounds are scaled respectively between [min,med] and [med,max] along the value of the interest. That way, the robot moves makes larger steps in low-interest areas.
This states is composed of several phases. First, we check on the current navigation goal.
- If it is reached, we set the next state to DETECTING.
- If it failed to reach it, we set the next state to PLANNING.
- If it is still going, we generate nextsteps new goals around the expected destination.
Then, if detect_while_moving is set to true, the robot will also run the DETECTING state.
If print_images is set to true, this node will output an image to the image_dir folder describing the current state of the probability field after every iteration in the DETECTING state. If detecting_while_moving is set to true, it will also output an image every moving_image_write_frequency detections while moving (as to avoid outputting several thousand frames). If simulate_source is true, the mean concentration field will be written as well, to image_dir/meanfield.png
These images can be compiled into a gif with the makegif.sh from neurovertex/infotaxis (though it requires image_dir to be "pictures").
Requirements:
- Compiler with C++11 support
- libpng and png++
- Catkin, obviously
Clone this repository in a catkin workspace and create a "lib" directory inside. Clone neurovertex/infotaxis somewhere, build it, then copy/move/symlink libinfotaxis.a to ros_infotaxis/lib/. From there, catkin build (or catkin_make) should work.
TODO