ROBOT SENSOR CALIBRATION VIA NEURAL NETWORK AND PARTICLE SWARM OPTIMIZATION ENHANCED WITH CROSSOVER AND MUTATION

In order to determine the position and orientation of an object in the wrist frame for robot, transform relation of hand-eye system should be estimated, which is described as rotational matrix and translational vector. A new approach integrating neural network and particle swarm optimization algorithm with crossover and mutation operation for robot sense calibration is proposed. First the neural network with rotational weight matrix is structured, where the weights are the elements of rotational part of homogeneous transform of the hand-eye system. Then the particle swarm optimization algorithm is integrated into the solving program, where the inertia weight factor and mutation probability are tuned self-adaptively according to the motion trajectory of particles in longitudinal direction and lateral direction. When the termination criterion is satisfied, the rotational matrix is obtained from the neural network's stable weights. Then the translational vector is solved, so the position and orientation of camera frame with respect to wrist frame is achieved. The proposed approach provides a new scheme for robot sense calibration with self-adaptive technique, which guarantees the orthogonality of solved rotational components of the homogeneous transform.Original Abstract: U cilju odredivanja polozaja i orijentacije nekog predmeta u zglobu za robot, treba procijeniti odnos transformacije sustava ruka-oko, sto se opisuje kao rotacijska matrica i vektor translacije. Predlaze se novi pristup koji integrira neuronsku mrezu i algoritam optimaizacije roja cestica s operacijom krizanja i mutacije za kalibraciju osjecaja robota. Najprije se strukturira neuronska mreza s matricom rotacijske tezine gdje su tezine elementi rotacijskog dijela homogenog prijenosa sustava ruka-oko. Tada se algoritam optimalizacije roja cestica integrira u program rjesavanja, gdje se faktori tezine inercije i vjerojatnosti mutacije sami podesavaju prema putanji gibanja cestica u longitudinalnom pravcu i lateralnom pravcu. Kad je zadovoljen kriterij terminacije, rotaciona matrica se dobiva iz nepromjenljivih tezina neuronske mreze. Tada se rjesava vektor translacije i postize se polozaj i orijentacija slike s kamere u odnosu na sliku sa zgloba. Predlozeni pristup pruza novu semu za kalibraciju robota tehnikom samo-adaptacije, sto garantira ortogonalnost rijesenih rotacijskih komponenti homogenog