A "Hand-Eye" Robot-Simulating System

Abstract

A "hand-eye" systea for simulating robots on BESI-6 computer is described. Peculia­ rities of the special operating system , problem-oriented software and hardware are discussed.

Elaborating of an "intelligent" ro­ bot is a versatile problem associated with the solution of various problems on information science, computing techniques, mechanics etc. The integrated investigati­ on requires a special experimental system for physical and digital simulation of robot functioning.

In physical simulation real sensors and effectors of robot are used the con­ trol and "intellectual" parts of robot being simulated by the computer. The al­ ternative kind of simulating system con­ sists in digital simulation both the ro­ bot elements and the surroundings. But in the latter case the detailed and reli­ able modelling of surroundings is very laboUr-consuming and found to be highly far from reality.

Our physical simulating system is developing as a "hand-eye" project. The hardware includea the computers BESM-6 and K-6000, the TV input-output device PTU-102, a special adapter with a storage tube, AD/DA conTerters, and a multijoint manipulator with six degrees of freedom. The manipulator was designed by Dr.A.Iv­ kin. The hardware includes also some addi­ tional apparata increasing the capacities of the BESK-6 computer for non-standard peripheral devices to be connected to.

The general-purpose software contains a special operating system having been worked out to provide the next facilities:

1. hierarchical structure of the pro­ blem-oriented software with mutually con­ trolled communication links between diffe­ rent levels of hierarchy and within one
2. level; 2) real-time and time-sharing mode of functioning;
3. direct man-machine communication with the simulating system;
4. special memory organization for processing of large volumes of informati­ on;
5. alternability and expandibility of the simulatin! system accordins to the tasks to be solved.

The operating system includes the available BESK-6 operating system DD-?3 plus a specially developed block '! ·\.lw.L·i-

G. Gimelfarb, E. Kushner, v. Rybak Institute of Cybernetics Ukrainian Acadeay of Sciences 25212? Kiev USSR

ver". The subdriver allows to form a complicated routine as a complex of relatively independent branches (up to 48), ensures multiprogramming mode of executing the branches and provides the time-sharing mode of information trans­ fer to and from the various non-standard peripherals. By means of the subdriver a programmer may synchronize the branches, change their priorities, exclude a branch when solving a definite task, etc. If no branch is ready to work the subdriver gives the processor up to the other job.

The subdriver is composed of highly autonomous blocks. It makes it ready to further expanding. The quasi-parallel and hierarchical mode of branch functioning are organized by means of a special addi­ tional set of control statements for in­ terrupting the processor.

The subdriver together with the spe­ cial auxiliary register of interruptions {1] allows to connect BESM-6 computer up to 48 non-standard peripherals. The maxi­ mum processing time for the one periphe­ ral interruption being only 80 sec. One can consider the operating system to be a real-time system where in~ut data call the corresponding branch [2J ·

Branches use a common memory that provides interchange of information. To overcome the difficulties involved with less then 32K BESI-6 core memory availa­ ble a dynamic loading of the branches from magnetic drum is now under develop­ ment.

It worth be noted that the operating system (OS) of the "Ryad" system gives many of facilities that the subdriver provides for BESM-6 computer. This ensu­ res a natural transference of this simu­ lating systtm from BESM-6 to the "Ryad" computer.

The problem-oriented software con­ tains various subroutines for input-out­ put, preprocessing, analysing the visuql and distance information and for control­ ling the multi-joint manipulator. The problem of automatic description of 3dimensional scene to draw up the purpose­ ful behaviour of the robot is of the grea­ test interest for us. The scene descrip­ tion includes recognizing objects and measuring their location and sizes. In our opinion there is a significant dis­ tinction between the problem or 3-dimen­ sional scene description by the robot and the problem of recognizing 3-dimensi­ onal objocta U5ing their 2-dime~bivuui images ( e.g. TV images). The distinction is caused both by the character of the information about the scene and by speci­ fic faculties of the robot.

The initial information for scene description includes besides the bright­ ness of image points also the values of spatial coordinates corresponding to some elements of objects. Specific abilities of the robot come from its capacity for active interaction with the surroundings. Being active system the robot can change the location and direction of the image sensor and light sources, remove objects and obstacles overlapping another objects with the help of its manipulator, measure the spatial coordinates of the different elements of the object, set parameters of sensors in accordance to results obtained during processing of the previously recei­ ved information, etc.

It is expedient to regard the senso­ ry system of the robot as a specialized system oriented on a specific class of objects. Knowledge of this class of obje­ cts and the robot abilities for the acti­ ve interaction with surroundings sympli­ fies the scene description problem and gives ground to contend that this problem is solvable in spite of comparatively low success in recognition of 2-dimensional arbitrary patterns.

One can note three basic features of the scene description process:

1. a complex usage of the informati­ on obtained from the different sensors ( image, space coordinates);
2. a hierarchical organization of the image processing with feed-back links between different levels;
3. a possibility to separate the who­ le scene description process into autono­ mous parts.

These parts are the following: a static scene description, a dynamic cor­ rection of the static scene description, a detection of dangerous situations. The first task in the most complicated one but after obtaining its solution the se­ cond and third tasks can be achieved in a real-time mode.

The multichannel sensory system of the robot permits to use simultaneously the information obtained from the diffe­ rent channels; in such a way a reliable scene description would be achieved in acceptable time and expenditures of com­ puter resources. For the present it is di1"ficult to point out the final cri teri­ on of using the each available channel. But temporary unability in choosing the optimal criterion should not be a stumb­ ling-stone for the research in multi-cha­ nnel sensory systems. The point is that an isolated utilization of channels often prevents to describe the scene obst:l.L"Vt:lu correctly.

'l'V image is the most informative channel but one can easily list examples of uncorrect scene description based on TV picture only. For instance, if small blocks are drawn on the visible sides of real block ( see Fig.1) the most part of known algorithms based on the edge extra­ ction will describe 4 blocks instead of the real one. The description of the lat­ ter will be found with the least confiden­ ce. The similar situation arises when si­ des of 3-dimensional object are pictured by a twisting lines forming camouflage. In such a case the edge detection is use­ less.

The traditional methods that used only 2-dimensional pictures to make the scene description meet considerable diffi­ culties even in more idealized cases when the scene containes only polyhedrons with single-coloured sides. To illustrate this words one can use the results of picture processing with the method being proposed by M. Suwa, s. Tsuji from ETL (Japan) {3] These results presented on Fig.2 demonstra­ te clearly all the complications in choo­ sing valid threcholds for contour extrac­ tion (Fig.2c), line thinning (Fig.2d) and tracing (Fig.2e), joining the neigbour line segments (Fig.2f), reconstruction the vertices and edges of sides (Fig.2g), etc. The subroutines implemented this method were worked up by us and included in our problem oriented software to be used for comparative investigation of the scene description algorithms.

From our viewpoint it is impossible to make scene description out from real noisy TV pictures without additional in­ formation about spatial coordinates of certain elements of 3-dimensional objects. One approach to receive this information leads to develop automatic methods of stereovision. We have investigated several problems in stereovision system hardware and software, some of the results being represented below.

The hardware of binocular stereovi­ sion system has no specific features to be marked out. The calculation links the values of system parameters with the cul­ minating accuracy in measuring spatial coordinates using a digitized stereo pair of TV images. The basic characteristics of the stereosystem were chosen in accor­ dance with the parameters of manipulator and manipulator jaw and are the following:

• interval of distance measuring: 1 to 2,5m;
• the least distance between optical axes of the stereochannels: 300mm; 2
• the least overlapping between right and left frames of the stereopair:
• 15mm;
• TV-picture frame: 24 x 30mm ;
• resolution of TV-picture digitizer:

20 points per mm;

• the least frontal 2 size of the ob­ ject observed: 50 x 50mm ;

- the least number of resolved points

for an object of the least size: 20 x 20;

• light diffusion circle : 0,05mm.

In this particular case one can show that an absolute accuracy of spatial co­ ordinate measuring needed to distinguish a decahedron from a octahedron and from a cylinder ought to be less then 1mm (i.e. a relative accuracy being less than 0,1%). 'rhe attainable relative accuracy of the system described is not less than 0,3%. Thus to achieve the accuracy requi­ red the stereovision system must be fur­ nished with the moveable image sensors to approach the objects observed and have changeable distance between the optical axes of channels. Using objectives of variable focus length only cannot solve the problem. The travelling image sensors can overcome difficulties associated with a high sensitivity of the binocular sys­ tem to turning angle of the polihedron· observed relative to the optical axes of the system (due to the different per­ ception of parts of the objects by the left and the right channel of the system).

The optical and geometrical design of the system takes into account the apparatus errors. Besides it there are abilities of the algorithms for automatic recognition of corresponding points on the stereo images that influence the accu­ racy of spatial measurements. The princi­ p~l fie~d of application of these algo­ r1thms 1s the photogrammetric automatiza­ tion (the problem of automatic reconstruc­ tion of the Earth relief). For the last problem we have proposed a few algorithms based on the assumptions about connection between the neighbour elements of the re­ lief and proportional changes of the brightness in the corresponding points at th~ lef t and the right images of stereo pa1r. Under these assumptions and som~ hypothesis about the characteristics of noise the problem can be solved by dyna­ mic programming (4] ·

The robot stereovision has its own ~eculiarities: the distance being measured 1s short enough to break the mutual con­ nection between the neighbour points of "relief"; the noise level is too high (due to noise of optical channels and image sensors). Thus an adaptation of the methods proposed to the robot stereovisi­ on becomes possible only on condition that one takes into consideration the content of the images compared. It means that the methods of fitting the corre·­ ponding points of stereo images should be com~i~ed with the techniques for the recogn1t1on of scene objects elements. Such combination of the different methods should simplify identifying the correspon­ ding points. One possible approach to such simplification consists in generating an artificial "texture" and illuminating the surface of objects being chosen du­ ring the recognition with the help of this "texture".

The calculations corroborates a practicability of this approach. For the present special subroutines are developed for computer investigation of the peculiarities of object surface illumination under various locations of objects and light sources and different reflectance of the surfaces. The subrou­ tines generates 2-dimensional images of the prescribed 3-dimensional scene for the various location and illumination conditions. A few examples of the such images are presented on Fig.3. The sub­ routines described are included in the problem-oriented software of the system.

The problem-oriented software con­ taines also a Guzman-like algorithm [5] fdr describing scenery composed of the ideal objects. An example of scene and result of the decomposition this scene obtained are shown on Fig.4.

An integrated algorithm for combi­ ned processing of the information from the different sensory channels of the robot is now under development.

We are grateful to Dr. I. Basilevich for his principal contribution to sub­ driver working out, and to our collabo­ r~tors A. Boldyrev, v. Gaiduk, G. Golo­ v~n, V. Krot, v. Lanbin for their par­ ticipation 1n the hardware and software elaboration.