This page contains a summary of ongoing and finished research projects in the field of intelligent robotics here at the department of Computing Science, Umeň University. If you have comments or questions regarding any of these projects, do not hesitate to contact the project leader Thomas Hellström.
Learn From Demonstration on Kompai Robot
The imitation technique is heavily utilized within robotics, in this case often denoted Learning From Demonstration (LFD) or Imitation Learning (IL). A human demonstrates a wanted behavior by tele-operating a robot. In this research, a robot will be equipped with a set of parameterized high-level behaviors. Models and techniques for ways in which behaviors can be combined into sequences will be developed. By recognizing a behavior on-the-fly during a demonstration, a shared control system can be constructed. In this way, the robot may take over control already during the demonstration. The human will be relieved of the hard task of tele-operating but may, if needed, interfere and correct the robot’s control signals.
System analysis using discrete event simulation
With continued development of forest machines the operators risk becoming a bottleneck, due to the ever increasing pace they have to work in. Different kind of aids for the operators will become necessary to lessen the workload. Development of autonomous forest machines aims at meeting this demand, and autonomous forwarders are considered to be a commercially profitable product scenario. The aim of this study is to validate this claim and check under what conditions autonomous forwarders could be profitable. This is done by using a discrete event simulator able to simulate the work flow, and thereby productivity, of future autonomous systems and comparing the results with the machines currently used in forestry.
This project is intended to develop software and hardware to be mounted on a full-sized forest machine. The goal is to make the machine travel without a driver along a predetermined path in a forest. Among others, the research includes algorithms for "path tracking", as well as detecting, locating, and avoiding obstacles.
More information is availible on the
A popular method for teaching robots simple behaviors involves a human demonstrating a behavior via remote control, or by having the robot observe the human's movements. In this project we endeavour to build robots that can learn new bahvior out of already learned or pre-programmed primitives. Teaching takes place by data mining of sensor data and control commands.
More information is available on the project's webpage.
USAR - Urban Search And Resque Robots
A model of a USAR robot that uses 3D camera and ultra sonic sonars to explore the environment and build a map. Obstacles are detected by transformation of 3D data to a floor coordinate system. Too high objects are shown in the user's display by overlaying color coded range data with a grayscale image. Simplified "humans" are detected and their locations memorized. Afterwards, the robot can be told to go to any of the found humans. A collision free path is planned using a watershed algorithm.
Check out the video
A Software Framework for Control and Sensing in Mobile Robotics
NAV2000 is a mobile-robot middleware implemented in Java, allowing efficient configuration of the robot's sensors and actuators. It supports, among others, the Pioneer and AmigoBot robots.
More information, and downloadable software are avalible on the project's webpage
Simulator in the loop for path planning
Many path planning algorithms make simplifying assumptions regarding the robot's geometry or kinematic behavior. Typically the robot is approximated by either a point or a circle, and the robot is assumed to be be able to turn in a predictable way given by simple kinematics equations. This usually works fine for regular indoor robots but is less suitable for large outdoor robots and autonomous vehicles.
This project is aimed at extending a standard path tracking algorithm with a simulator that, in real-time, tries to predict collisions in a window forward in time. This simulation is based on current sensor data giving information about the environment around the vehicle. If a collision is predicted to occur, a path optimization phase is initiated. Variants of the original path is generated and simulated until a feasible path is found. The real vehicle may then continue, now tracking the replanned path.
More information can be found in our journal publication.
Most of current robotic applications are based on the "Classical AI" paradigm. Localisation, Computer Vision, Map Building, Planing, are techniques used to solve a large range of real-world applications. Finding, cutting, and gathering timberwoods belong to them. But which kind of robot could climb a tree and attach a cable on the top of it?
On this kind of problems, Classical AI reveals its limitations. Indeed, it has more to do with agility, real-time adaptivity, and embodied / animal-like intelligence. This project adresses this issue through Artificial Evolution.
More information is available on the project's webpage.
Localization by using a laser scanner
We analyze images from a laser scanner, and determine the
shift (position and direction) by comparing two images. The technology
can be used for relative localization ("laser odometry") by calculating
relative movements between two pictures taken in sequence, and also for
absolute localization by comparing a recent picture with pictures from
A miniature robot equipped with a camera is used for illustrating the
idea of a cleaning robot for the forest industry. Our mini-robot operates
in-house and can distinguish between various trees and different log shapes,
using histogram analyses of snapshots and a small database of various
tree types. More information is available on the project's
Neural visual odometry
This project is intended to develop technologies for
determining a forest vehicle's movements, using cameras pointing
downwards to the ground. Techniques for correlating pictures and
optical-flow algorithms are used to determine the pose change between
two consecutive pictures.
In this project we investigate sensor technologies to detect rocks in mines.
Boulders often present a problem when transporting ores in LKAB's subterranean
mine shafts. By detecting and then knocking down the boulder, costly work
breaks can be avoided.
Caro - the Candy Robot
KiKS - Khepera Simulator
KiKS is an abbreviation for "KiKS is a Khepera Simulator". As the name suggests, KiKS simulates a Khepera robot connected to the computer in a very realistic way and runs under Matlab. You control the simulated khepera using three commands (kiks_kopen, kiks_ksend, kiks_kclose) which work just like the K-team kMatlab modules (kopen.dll, ksend.dll, kclose.dll), so if you are familiar with controlling the khepera robot from Matlab you can instantly start using this simulator. Also, creating programs that can run on both simulated kheperas and real kheperas is extremely easy since the kiks_kopen/kiks_ksend/kiks_close commands work as wrappers for the kMatlab modules - calling kiks_kopen with a negative port number opens up communication with a simulated khepera, and calling kiks_kopen with a positive port number simply redirects the call to kopen.dll.
KiKS started out as my master thesis and has since then been further developed. More information can be found here.