Embedded Files
This project was development within the "Young Teams grant: Reinforcement learning and planning for large-scale systems" grant (link).
Our goal is to use a mobile robots in a home environment to help elderly or disabled people in their daily tasks. A household is a changing environment. Thus, we model it with the Partially Observable Markov Decision Process (POMDP). The project is composed of two main parts: analyses of the applied method and implementation. I worked partially on the analyses part, and details can be found here, and substantially on the implementation.
Details about the application and the implementation
Details about the application and the implementation
Our mobile robot system is composed of a Pioneer3AT mobile base, extended with a Sicklms200 laser scanner, and on top of it is mounted a Cyton Gamma 1500 robot arm. The problem setup is a real-life test for the POMDP solver, so we simplified some aspects that are not relevant for the evaluation.
Scenario and tasks: We defined a "simple" real-life scenario with the necessary characteristics to be modeled with POMDPs. The goal is to keep turned off some switches in a known environment. The robot position is always known, while the state of a switch might not be observable from a certain distance or position due to perception noise and occlusion, respectively. Thus the system can be modeled with a Mixed Observable MDP (MOMDP) due to the fully observable (robot pose) and a partially observable variable (switch state).
The MOMDP model:
Robotic arm: I was highly involved with the switch state manipulation problem, where we used a low-cost webcam mounted above the gripper. The arm was controlled with the MoveIt ROS package and a visual perception packages to close the control loop.
The arm's controller (ROS package) can be found here.
Rviz view of the robot and the detected swich
Some early demonstrations about the navigation and flipping a switch :
Implementation of the solver integrated into the SARSOP framework
Implementation of the solver integrated into the SARSOP framework
GIT repo of the entire system can be found here: https://bitbucket.org/ElodP/home-assistance-application/overview
Reference publication: Elod Páll, Levente Tamás, Lucian Busoniu, Analysis and a Home Assistance Application of Online AEMS2 Planning. Accepted at IEEE/RSJ International Conference on Intelligence Robots and Systems (IROS-16), 2016
Google Sites
Report abuse