John Folkesson

Movies:
• Here  is Pluto making a large outdoor map using a Compressed Extended Kalman Filterd Compressed Kalman Filter.   The robot is the small x.  The map is a surveyed map of the buildings and very accurate.  The walls are detected by a SICK laser scanner and formed into a map.  Each new measurement is used to improve the map and the location of the small x.

• I have improved the estimate of the maps and location by developing a graphical SLAM method.   Here is an example of using Graphical (Robust) SLAM  to close a large loop, avi: Graphical Slam With Loop Closing .  The dots being left behind the robot as it moves are additional states being estimated.  Having these additional states allows better linearization of the system.    It also creates a sparse structure of connections between the features that can be exploited to speed the calculations.  After this work was done there was a wave of interest in sparse SLAM methods.

•  Here is a movie showing a robot mapping  our lab using a camera (slowed down to half actual speed of  the offline calculation) EKF SLAM using Vision  and M Space.   This uses a standard SLAM but using a camera to do SLAM with constraind 3D lines is tricky.  We developed a way to represent the constraints on the lines so that they could be used as 1 or 3 dimensional features.

• Here is a movie showing a map of our lab using a camera and SICK scanner, EKF SLAM using Vision, SICK,  and M Space. Using different types of sensors presents a challenge in fusing the information from the sensors that can be qualitatively different.

• Here is a movie showing a graphical map of our lab using a camera, Graphical SLAM using Vision and M Space.

• Here is another  movie showing a graphical map of our lab using a camera where we merge star nodes  to build constraints into the graph, Graphical SLAM using Vision and M Space.
• Here is a movie showing a graphical map of our old lab.  Here we detect  3 separate loops automatically and enforce the constraint on the graph. Closing the Loop Automatically with Graphical SLAM.
• Something completely different, the   Antiparticle filter. Compared to the   Particle filter. Here we simulate how very noisy odometry can be corrected with proper correllations in an adaptive analytic recursive Baysian filter.

•  Moving Underwater

• Here is a movie showing  an autonomous underwater vehicle, AUV, matching a SLAM map to an a priori map: Underwater SLAM.  This is a complete navigation system.  It includes 5 estimators: 3 EKF's, a prediction filter, and a tracking filter.  The tracking filter uses a novel feature representaion and a graphical SLAM algorithm to make sense of sonar and motion data.  It passes what it learns as chunks of information on entire local areas to a global EKF.  This happens at a slow rate (every 15-100 seconds).  The global estimator can then match to the a priori map.  This system has since  developed into a very robust navigation system that can handle large amounts of ambiguity (see below).  It is fast enough to be used for control, (and we do that).
• Here is another  movie showing  the  AUV matching a SLAM map to the a priori map.  The  correct matching is chosen at last but the robot chooses some more aggressive hypothesis first.  Then, as more information comes in, it switches to a more likely conservative matching hypothesis: Multihypothesis Matching.  Here is a simple matching that looks to a human like it works but not if you look at the first movie; this was the live result: Almost Right.

• We went back and re-thought things and two months later... Here we have a perfect day of testing.  A rare event in robotics.  The goal here is to have the AUV find the designated target (the circled star in the middle of things).  The robot must match the confusing sonar pings in light blue with the approximate a priori map, the  tiny purple dots.  The double rings are matched features: Five for Five.   This uses a new matching metric that we plan on publishing a paper on soon.  Here is a longer look at the forth run that day  Run4
• We have since added an attachement mechanism to the robot and done ocean trials with high success rates in attaching to the chosen target.  Here the moored target is at 20 feet altitude and the last in the third line.  The robot uses the bottom features to navigate to the target then increases its altitude before attaching to the line.

Contact details
John Folkesson.
Kungl Tekniska Högskolan
SE-100 44 Stockholm
Sweden

 Email johnf@csc.kth.se

Generated by John Folkesson, johnf@csc.kth.seDecember 17, 2010