Mapping and Localisation
Almost all applications of field robotics require the robot to be able to locate itself with respect to some coordinate system. Our research and expertise in this area includes laser scan matching, beacon-based localisation, visual odometry, and off-board vision-based localisation (using fixed cameras to track the robot).
Another common scenario in field robotics occurs when a robot has no known map or accurate measure of its pose, so must simultaneously estimate and maintain both. This is known as Simultaneous Localisation and Mapping (SLAM). Our research focuses on solving the SLAM problem for applications involving long term autonomous operation of robotic vehicles.
Research Topics
 |
 |
 |
 |
- Much of the experimental data illustrated below has been collected at the QCAT site near Brisbane, which offers a variety of environments including grassed areas, roadways between buildings and trees, and an industrial compound.
Simultaneous Localisation and Mapping (SLAM)
Robots that operate in environments where no a priori map is available must often construct one in order to function and adapt to their working environment, or to provide accurate models for an end user. Simultaneous Localisation and Mapping (SLAM) is a fundamental problem in mobile robotics, describing the common scenario in which a robot with no known map or accurate measure of its pose must estimate and maintain both.
The SLAM problem is made even more challenging when long-term autonomous operations require that a robotic vehicle repeatedly re-traverse an environment. A requirement for feasible operations in these conditions is that the mapping algorithm does not require an unbounded amount of memory. Therefore, redundant local maps need to be detected and consolidated in the global map. Another necessity for long-term operations is the ability to recover from intermittent sensor outages which may see a vehicle in a completely different location when its sensors come back online. The system must therefore be able to perform a global re-localisation when the robot becomes lost in this way.
- Our place recognition algorithm enables the generation of accurate 2D maps in real-time. In this example, a metric street map is constructed using data collected from two lasers mounted horizontally on the roof of a vehicle. The data was collected while driving at road speeds (up to 90km/h) for over a total of 165km throughout Brisbane. The colours indicate individual local point cloud maps, which are globally aligned using the overlaps identified by the place recognition algorithm. The Brisbane CBD can be seen in the upper right of the map, while QCAT is located in the bottom left. Each grid cell is 1km x 1km.
We have developed a technique for Simultaneous Localisation and Mapping utilising 2D laser range data that does not require input from other odometry, inertial, or GPS sensors. The laser data can be used to build local maps using a robust incremental scan-matching algorithm. The local maps can then be globally aligned by finding adjacencies using a place recognition algorithm and a series of map matching techniques. Using this approach, accurate and consistent maps have been constructed from hundreds of kilometers of vehicle trajectory. The place recognition technique can also be used for localisation in known environments if a prior map is available, or to merge maps collected at different times or by multiple vehicles.
We have further enhanced this technique to construct 3D maps by mounting a commercial 2D laser range finder on a spinning platform which is then fixed to the roof of a vehicle. An additional challenge in the 3D scenario is that collecting consistent 3D laser data on a moving vehicle is problematic, as the scanning time can be significant relative to the motion of the platform resulting in severe distortions of the uncorrected data. In order to solve this problem we have developed a 3D sweep-matching algorithm that estimates the trajectory of the sensor while it scans without requiring the use of any additional sensors. In this way we are able to use relatively inexpensive sensors to estimate motion and generate locally consistent 3D maps as our vehicle continuously navigates. By applying our place recognition algorithm, we are able to reliably close loops to create globally consistent maps.
(See Publications below for more information on this research.)
- 3D map of the QCAT site built up using a commercial 2D laser mounted on a spinning platform on a vehicle. Buildings and trees can be made out along the path of the vehicle.
- Detailed 3D map of the compound area.
We are interested in the problem of automating industrial vehicles for operation both indoors (for example in warehouses, smelters, hangars, etc) and outdoors. A primary requirement for such vehicles is an accurate localisation system to support both navigation and tasks such as moving material or equipment.
- Laser sensors fitted to the Hot Metal Carrier.
GPS is not a suitable technology for this application due to its limited accuracy and unavailability indoors. Additionally, the long-term accuracy of maps or techniques such as SLAM can be degraded by changes in the environment resulting from the activities of other vehicles, the placement of materials (pallets, containers, etc), or the growth of foliage over landmarks. Alternative localisation methods include placing magnets along the path the vehicle is expected to travel or placing visual markers (detectable by an onboard camera) around the environment. However neither of these systems is suitable in applications where the autonomous vehicle is required to travel outdoors around a site containing large open areas.
In order to provide a reliable set of landmarks in such an environment we have placed laser-reflective beacons at surveyed locations about the site, thus establishing globally-accurate locations. Our vehicle localisation system then uses onboard movement sensors and laser rangefinders to calculate its pose (position and heading) using a particle filter.
We have employed this localisation system on a hot metal carrier vehicle both at our industrial work site and in a trial deployment in an operating aluminium smelter. The localiser has demonstrated position accuracies of up to 5cm.
The increased popularity of wireless networks has enabled the development of localisation techniques that rely on WiFi signal strength. Such systems are cheap, effective, and require no modifications to the environment. Although there are many examples reported in the literature of such systems being used for person tracking and location-aware computing, it is unclear whether a WiFi localisation system alone will ever provide the accuracy required for autonomous vehicles operating in industrial environments. However, as a secondary localisation system, WiFi signal strength information can be used to generate a coarse global position estimate which can be used to bootstrap the a vehicle's primary localisation system. We utilise an existing 802.11 wireless infrastructure to estimate the location of an autonomous vehicle within an industrial environment.
We have designed the WiFi position estimator for use in one of several scenarios. Firstly, it can be used as as a prior estimate during the initialisation of a primary localisation system. This can speed up early convergence and reduce the possibility of multiple hypotheses due to environment symmetries. Secondly, the WiFi localiser can be polled when the primary localiser's confidence drops below a threshold, indicating that it may be lost. In this case, the WiFi localiser can again provide a position estimate to re-seed the local position estimator. Finally, the estimate from the WiFi localiser can be used as an independent estimate to detect possible anomalies occurring in other localisation systems.
Our approach to WiFi localisation is tailored to industrial environments. The two main components of the system are a WiFi map generated using Gaussian processes, and a localiser which uses Bayesian filtering. The resulting system was analysed against ground truth measurements to determine the expected accuracy which found it to provide acceptable and consistent accuracy for its intended usage. To demonstrate its bootstrapping ability, successful experiments were conducted using the WiFi localiser to provide an estimate used during after triggering a failure in the primary localiser.
- This illustration shows a map of the work site with WiFi access point locations, and different environments (compound, building canyon, and open roadway). Also shown are the localisation accuracies for each of the areas.
- Starbug, CSIRO's autonomous submarine. When equipped with an acoustic node it is able to track its position relative to an underwater sensor network using a particle filter.
Autonomous Underwater Vehicles (AUVs) operate without human control or supervision, performing various duties such as the inspection of natural habitats and marine infrastructure. Unlike other types of field robots, AUVs have significant challenges determining their location as GPS signals are not available while underwater. Tracking the AUV's position is important for navigation and for determining the world location of the data that it gathers.
The ICT Centre has been developing an AUV localisation system based on underwater sensor networks which use acoustic modems to transmit data between sensor nodes as sound waves. When equipped with one of these modems an AUV is able to communicate with the sensor nodes and, from the round trip time, determine how far away the nodes are. By integrating these distance measurements the vehicle's position can be determined.
|
|
Key Staff
Publications
- Michael Bosse and Robert Zlot. "Keypoint design and evaluation for place recognition in 2D lidar maps." Robotics and Autonomous Systems. Volume 57, Issue 12, 31 December 2009, Pages 1211-1224.
- Michael Bosse and Robert Zlot, "Place Recognition using Regional Point Descriptors for 3D Mapping", International Conference on Field and Service Robotics, Cambridge, 2009.
- David Prasser and Matthew Dunbabin. "Sensor Network Based AUV Localisation," International Conference on Field and Service Robotics, Cambridge, 2009.
- Michael Bosse and Robert Zlot, "Continuous 3D Scan-Matching with a 2D Spinning Laser", IEEE International Conference on Robotics and Automation, 2009.
- Felix Duvallet and Ashley Tews. "WiFi position estimation in industrial environments using Gaussian processes". IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS 2008); Nice, France. IEEE; 2008: 2216-2221. ISBN: 9781424420575.
- Frederic Moster and Ashley Tews. "Practical Wifi Loalization for Autonomous Industrial Vehicles", Australian Conference on Robotics and Automation, 2006. (Best student paper award).
- Robert Zlot and Michael Bosse, "Place Recognition using Keypoint Similarities in 2D Lidar Maps", International Symposium on Experimental Robotics, July, 2008.
- Michael Bosse and Robert Zlot, "Map Matching and Data Association for Large-Scale 2D Laser-Based SLAM", International Journal of Robotics Research, 27(6), June, 2008.
