A Bioinspired Neural Network for RealTime Concurrent Map

A Bioinspired Neural Network for Real-Time Concurrent Map Building and Complete Coverage Robot Navigation in Unknown Environments Harry Potter and Miika Ranimäki 4. 2. 2019

Cleaning robot • Cleaning work falls into the category of “dirty, dull, and dangerous (3 d)” • Great market demand • Cover whole region, minimal turns, minimal overlap • Map building and collision-free navigation important

Proposition • Complete coverage navigation (CCN) = going through whole map • a novel neural-dynamics-based approach is proposed for real-time map building and CCN of autonomous mobile robots in a completely unknown environment. • Most neural-network based models require learning procedures which are computationally expensive and difficult to achieve CCN in real time

Environment and Map • Dynamic environment • Completely unknown • 2 -D Cartesian workspace • assumed to be known • Consists of rectangles

The Model • Biological neural system using electrical circuit elements • Robot motion is planned in real time based on the dynamic activity landscape of the neural network and the previous robot position, to guarantee all areas to be cleaned and the robot to travel a smooth, continuous path with less turning. • The model is extremely complex and works utilizing a three-layered neuron network

Coverage and map-building algorithms • Each neuron is labelled unknown, cleaned, obstacle or deadlock (Imagine 2 D array[k, l]=label ) • Cleaning robot can ONLY know workplace size, nothing of the neurons • The map building relies on robot’s onboard sensors and is done dynamically • After the map is built starts the CCN phase, where the robot is attracted to unclean areas and continues to clean until the workspace is fully cleaned

Fully known environment

Unknown environment

Dynamically changing environment

Algorithm • Initialization • set all areas as unclean • Set all neural activities to zero • Coverage • • • Scan unknown neighboring neurons Compute neural activity Find the next neighboring neuron with the maximal neural activity Set current neuron to neighboring neuron if neighboring neural activity <= current neural activity: mark as cleaned repeat

Triangular cell decomposition approach Conclusion • The proposed model is effective and economic compared to competition • Works in completely unknown environment and needs no learning procedures Spanning tree approach

A Weight-based Map Matching Method in Moving Objects Databases Miika Ranimäki 4. 2. 2019

Weight-based Map Matching Method • Solves Offline snapping • Can get up to 94% correctness • Uses location management format (x, y, t)

Offline snapping • Finds overall route that was taken after the trip is over • Straightforward methods just use the output of GPS receiver and snap it to road without weighting, leading to imperfect results • Typical GPS error range up to 10 meters • Do not discuss the role of the time interval between two consecutive GPS samples

Variables • Tr = Trajectory from A to B • R = Snapped route • Similarity of Tr and R = sum of the distances between Tr and every arc of R • M = Map • e = arcs in M • Smallest weight path between start e and e is R of Tr

Method • Tr = Trajectory from A to B • R = Snapped route • Similarity of Tr and R = sum of the distances between Tr and every arc of R • M = Map • e = arcs in M • Smallest weight path between start e and e is R of Tr • 3 D weight algorithm: the motion of the snapped route should be close to the motion of the trajectory in 3 D • 2 D polyline can be raised to 3 D by using the timepoints ti and tj of start e and e in linear interpolation. Then the weight is defined as the integral of Euclidean distance between the subtrajectory from ti to tj, and the 3 D arc, divided by |tj – ti|. • The Tr is calculated by using Dijkstras algorithm with these weighted e: s

Conclusion • A Weight-based Map Matching Method in Moving Objects Databases solves Offline snapping by assigning arcs with weights • Much more accurate than straightforward methods • By itself this isn’t very useful, but the information could be used in the future for optimizing pathfinding.