Researcher in Autonomous Sensor Systems
Note: While the above figure obviously depicts a batch variant of the basic algorithm, this is purely for visualization purposes – this is in fact a truly on-line, incremental approach to SLAM, as reported in our J. Autonomous Robots article “Fast Online Learning of Globally Consistent Maps”. Our approach takes gradient descent steps by moving a node at a time (“pick a node and move it to where its neighbours think it should be”). Each full iteration of the algorithm updates every node (in a fixed order, though this could also be random). An incremental state representation allows new poses and new constraints to be added without restarting the optimization. This enabled a fully autonomous exploration strategy for map learning where the current state of the map was used by the robot to decide where to go next (so that we did not need to drive the robot by hand or work off-line with pre-recorded sensor data). This is also reported in my PhD thesis. Later work by Frese et al. extended the approach using multi-grid optimization techniques. There are some further subtle and important differences between the different algorithm variants, though this should not confuse the fact that all of our work was developed with incremental, on-line use in mind (29 Aug 07).