Follow me

Wednesday, October 3, 2012

A hidden Markov model-based assembly contact recognition system


H.Y.K. Lau
Mechatronics, No.13, 2003
Summary
                  Robotic assembly is a defined as a classical problem in industrial robotics where a rigid robot with accurate actuators performed sequences of predetermined operations in an assigned workspace through a position-based controller. Limitations during assembly tasks appear to be due to fixtures, features and tolerances, in order to overcome these limitations force-based control strategies have been proposed, such as: spring damper, hybrid force/position control, impendence control and so on. The fundamental issue in robotic assembly is the ability to recognize assembly states and this should be addressed before the deployment of appropriate control strategies. The proposed system (HMM-based contact state recognized) can be considered a high-level feedback control system for perceiving the environment in terms of symbolic expressions. With Fundamental Contact (FC) the author intends to describe contact formations that may consist of zero, one or more pairs between the work piece and the environment in which it is involved (it is a primitive for of contact formation involving a single pair of contact features). A contact is defined as a result of geometric arrangement, contact based control strategy for robot assembly is based on the pattern of force/torque which is formed between a work piece and the manipulator, so that the d.o.f. is reduced at least by one. Different contact states are then defined: 1) Configuration of an object (c, is the set of kinematics parameters that locates an object for a given set of joint angles); 2) Configuration Space (C, the Rn space which collects different c); 3) Contact Feature ( F{f,e,v}, where f is the face, e the edge and v the vertex of the object); 4) Fundamental Contact (FC ordered pair of contact feature; 5) Contact State (CS is a set of FC between to polyhedra).
In robotic assembly Contact Recognition is defined as the process using a contact sensing techniques to obtain geometrical information of contact with corresponding symbolic interpretations. Symbolic interpretations of an assembly provide high-level knowledge of assembly operations to robot programmers, but some limitations in the models (Petri-nets, ANN, Rule-Based contact analysis and further on) have been encountered due to complexity. HMM, already used successfully for speech recognition, is proposed considering assembly states interrelated and occurrence of contact formation between a work piece and the mating parts normally distributed.
·       Model
A HMM is composed of two stochastic processes, a basic Markov chain and an observable stochastic process. HMM are defined with a triplet λ=(A,B,π) (A is the transition matrix for the transition probabilities, B the matrix of probabilities for the observables and π is the initial state of the distribuition). HMM can be used to generate the observation sequence, mainly observation probabilities can be evaluated, the most probable sequence, given the observation sequence, can be also obtained and the probability of a sequence given an observable for a certain triplet can be maximized to find the HMM performance. The architecture of the HMM-based contact state recognition works in the following manner: defining the force torque signals, process them, map them and transfer the data to the HMM, after which the model with maximum probability must be selected in order to finally classify the state. The method, which for its final implementation uses forward and backward algorithm, Baum Welch algorithm (Appendix A),ALBG-VQ algorithm and symbol mapping algorithm (Appendix B), has been tested to investigate sensitivity, performance of recognizing 2D and 3D contact formations and the ability to classify sequence of contact formation during peg-in-hole tasks.
Key Concepts
HMM, assembly contact recognition
Key Results
The system is insensitive to the number of states and small amount of training data is required. Superiority of HMM-based system has been shown compared with other traditional methods already mentioned above.

No comments: