Neural network for black-box fusion of underwater robot localization under unmodeled noise
- 1. Neural network for black-box fusion of underwater robot localization under unmodeled noise 19 sept 2019
- 2. Types of underwater vehicle • Autonomus Underwater Vehicle (AUV) with minor or no intervention from operator • Remotely Operated Vehicle (ROV) for maintainance, repair operations and sea inspection
- 3. undewater environment and sensors • rapid attenuation of higher frequency signals and unstructured nature of water • for postion and movement acoustic, vision , accelerometer and gyroscope sensors are used
- 4. need of research • there should be a neural technique for sensory data fusion • coherently combining of data(any) to estimation location with respect to reference frame
- 5. Acoustic sensors comparision 1/2 • cost functions (rate, reliability etc) • acustonic positioning sensors provide asynchronous mesurements • estimates do not drift over time in acoustic sensors so long run measurement more reliable • slow signal traveling rate (1500m/s)
- 6. acoustic sensory comparison 2/2 • geo refered landmark are more reliable than acoustic sensor • commonly accoustic (easy) LVS example • robot have to come to surface to reduce localization uncertainty
- 7. main problem of fusion • baysian algorithms(kalman filters) are fine but estimation perform poorly under unmodeled noise • parametric algorithms (MCL) do multimodal hypothesis with high computational cost • problem is to make unimodel estimates under unmodeled noise
- 8. Approches for Problem 's solution 1/2 • model non linearity by agumenting state representation (difficult) • supervied learning methodologies in training of fusion algorithm for correcting estimates (alternative) {but for supervised learning task condition should not vary in training and execution time}
- 9. Approches for Problem 's solution 2/2 • Fusion of redundant estimates of each sensor based on error covarience matrix (inversly or covarience intersaction) • fusion process as fuzzy rule based system (home example of neural network training!!!!)
- 10. Writers' research focus • developing heuristic and generic fusion policy for redundant estimates to handle unmodeled noise • proposed architecture where redundant estimates are viewed as black box processes
- 11. Information fusion (past work same writer) • principle of contexual information anticipation for obtaining more reliable fusion for egocentric localization • reliability is evalutated within processing neighbourhood (mean and deviation) • confidence is in context of task and nodes contrubution accordingly weighted
- 12. Theory - Fusion architecture • blue node has fusion algorithm and weights for each estimator • Reset feature (blue node) for reducing error (dead recking replace with global estimate) and context transition according to task
- 13. thory- fusion arhitecture- Parameter set • µi(t) is mean state estimate • Σi(t) covarience matrix related to error • σi(t) is expected deviation of neighbor node from mean • δti(t) is time interval between two esimates • i is process and t is time
- 14. theory - fusion architecture - ordering arrangements1/2 • fusion node's ordering rule depends on two assumption • A1: Estimates from non delayed measurements and follow unimodel distribution(mean and covariance matrix)
- 15. theory - fusion architecture - ordering arrangements 2/2 • A2: reliability in relation to behaviour profile (antisymetry, transitivity and totality)
- 16. Theory - Modeling of A1 and A2 • assume set of one dimentional site arrangement (under B profile) • Sb is neighborhood system • relationship properties: -no neighboring to itself -neighboring relationship is mutual clinque left and right
- 17. Theory- Purpose of neural network • To model arrangements under distict behaviour profile • weighting information from redundant estimates to fusion process
- 18. B-PR-F for neighboring arrangements • Layer B is for behaviour • Layer P is for prediction • Layer R is for reliablity • Layer F is for Fusion
- 19. B-PR-F layer B • Task senerios and conditions(near surface or seabed) represented by behavior profile • cardinality is determined by k availble behaviour profiles
- 20. B-PR-F Layer P (imp noise) • contextual anticipation b/w neighborhood arrangements • neurons here encode parameter σi(t) • σi(t) is expected deviation from sorrounding µi(t) where estimate of process i should fall • activation becomes stronger and uncertainty increasess as task progesses (with motion of robot and time pass)
- 21. B-PR-F Layer R • encode confidence on nodes's estimate in relation to predicted value • node passes the test and related node is re-initialized • cardinality is of PR layers depends upon Behaviour and estimators
- 22. B-PR-F Layer F • contains fusion weight for estimaters • cardinality is determined by n number of estimaters • global estimated is calculated by these weighted sum
- 23. B-PR-F - Network parameters • Wbp condition the activity of PR layer according to B profile • Wpp represent lateral connection of layer P to model the changes(due to B profile) -no Interaction then Wpp is identity otherwise -f(arbitary func) is defined according to task and estimator parameters
- 24. B-PR-F - Network parameters • Wpr connectivity between neighborhood • ς is high magnitude assigned to unrelated neighbors (inhibation) • Wy is ordered arrangements of nodes(according to B profile) analytically stronger weights to clinque , excitatory weights to left and inhibitory weights to right neighbor • Wℵ(kxn) is provided by system designer which tells reliability along behavior profile system configuration matrix difference tells ordered arrangements
- 25. B-PR-F Network parameters • Wrf obtains from row of system desinger matrix • correspondance b/w sites established through maping function (represented in PR and estimators)
- 26. B-PR-F Layers(Behaviour) activation • Winner takes all policy • only one behavior at a time • arbitary function (self defined that behaviour is 1 or 0)
- 27. B-PR-F Layers(Prediction) activation • Activation of neurons in layer P • reset function cosider reliability test applied to right neighbor • time scaling factor is representing by heuristic parameter reset function time change heuristic parameter ** P node contains deviation
- 28. B-PR-F Layers(Reliability) activation • information is determined by left neighbor's parameters • activation of layer R indication of new estimator(given instant) threshold value obtained from left estimator
- 29. B-PR-F Layers(fusion) activation • weighting the n estimators’ output proportionally to the activation of Layer F (for u and Σ )
- 30. Simulations - Materials and methods • simulation in Gazebo • oceans waves senerio in robot operating system • dataset is of way point trajectory • robot to pass near way points with particle physics engine(to model dynamics from equatic medium) • Data generated is passed to GNU
- 31. Simulation - sensory details • Predefined trajectory with exploration mission of shallow and deep water • IMU for linear acceleration and rotational rate(surface) • DVL is more sophisticated in deep water • USBL and DGPS are positioning sensors (in deep water only usbl)
- 32. simulation results • ALT for switching Behavior profile • IMU for near surface and DVL for Deep water produce precise result • USBL can face issues (like physical interference) • BPRF eliminate noise of virtual usbl sensor
- 33. Simulation results (behaviour 2) • Principle of contextual anticipation within ordered neighborhood (only XY axis) • boundary of anticipated region { r = p6(t)Σ2(t) i } encode deviation of E3 with respect to E2 estimate
- 34. simulation - Activation of layers • evolution of layers for above anticipated trajectory • if usbl is within anticipated region then reset signal for coresponding node of P(anticipation in stand deviation unit) • neighbor arrangements and evolution of info is encoded by network • F can reject unexpected disturbance
- 35. simulation - activation of layers - lateral connection role • Wpp is set according to two different condition (identity and behavior change) • define sparse encoding matrix (which maps Behaviour a units to Behviour z's)
- 36. Experience • travelled 396 in 61 minutes • Sonar tilt is 0 regarding horizon • it can cover 130 horizon and 50 meters • 19992 grayscale 16-bits images by the SONAR • 3663 leading values of compass • 3662 positions by DGPS • 1450 by USBL
- 37. Experimental Fusion of sensor
- 38. Experimental - localization - scan matching motion estimation • relative motion is from multibeam sonar(no need of correction) • parameter of interest • dead recking is obtained by integrating relative displacement over time
- 39. Experiment - Kalman filter • F is state transition model • x is estimated state vector • B is control input model • u is control control vector • e is random gaussian vector that model uncertainties introduced by state transition • posterior state is corrected by x P y S K predicted state estimate and covariance