The αPOMDP framework for User-Adaptive Decision Making
This package contains two main elements:
- Our simulation suite (Julia code used in our simulations, Python plotter);
- A ROS package for interfacing with real robots.
The whole code is packaged as a ROS package, with both ROS and non-ROS-specific code in it. This allowed us to have a single implementation of the aPOMDP "class" that is used both by the simulator and the ROS stuff.
The package is composed of a few main files:
apomdp.jldefines the aPOMDP object, which is the core of our solution. It contains our definitions of the POMDP problem, including our learning facilities and the various rewards we have implemented. It also contains a function to log the results of a trial run.
plotter.pyis a python script that can be used to plot the results of any trial in which the logging function is used. It produces the figures we have used throughout the paper.
apomdp_tests.jlcontains a number of small "unit tests" for the aPOMDP object that come in handy to check if everything is running smoothly.
hri_simulator.jlimplements our simulator. It basically consists of a main function that implements the interaction loop (get action, get reward, transition state, learn), with a few auxiliary functions.
policy_calculator.jlis our ROS-facing interface for
apomdp.jl. It serves a ROS service defined in
GetAction.srv, wherein it receives an observed state and returns the appropriate action. It simultaneously (and in parallel with the service) integrates the observed transitions for learning, and re-calculates the policy after each integration.
apomdp_conductor.jlimplements similar logic as the
hri_simulator.jl, interfacing with the GrowMu robot in order to speak, move around, listen to the user, etc...
This package was tested under:
- Julia 0.6 (0.5 is known not to work)
- ROS Indigo (later versions should work)
- Latest versions of JuliaPOMDP and related packages, as of October 2017
- ROS interfacing was obtained using the RobotOS Julia package
The main entry point is the
hri_simulator.jl script, which is our main simulator. This makes use of the "class" defined in
apomdp.jl. The script contains the various configurations used for producing the results in the paper.
To use this package with a robot, you will need to use the
apomdp_conductor.py script as a guide for implementing your own script that interfaces with your robot. In our case, we have used the GrowMu robot, and our proprietary (maybe I'll get to release it some day) toolset for interfacing with it.
You will also have to configure the
policy_calculator.jl script to work on your definitions of the action and observation/state spaces. These should be a matter of re-configuring the parameters used to construct the
In both cases, YAML files containing the results will be produced. The
plotter.py script can be used to obtain the illustrations used throughout the papers from both data sources, simply by configuring which files it looks at in the main function.