• Contribute
• What's New
• Submit News

Robots everywhere

Robots. Until recent years hearing this word we could think about only one type of existing application: industrial multi-joint systems working at assembly lines, welding car components, or sorting soda cans. The most important task of these automats is the reliable, long term work.

Figure 1: Welding robot (Kuka Robotics) ([1])

Figure 2: FlexPicker multifunction robot arm ([2])

In the last two decades a new category of robotics, the mobile robots evolved significantly, so nowadays they reached the level of everyday usage.

Figure 3: Lawnmower robot ([3])

Figure 4: Vacuum cleaner robot (Karcher) ([4])

These facts are supported by the World Robotics 2004 survey of United Nations Economic Commission for Europe: globally there are one million industrial robot in service and the number of household robots at the end of 2003 hit six hundred thousand ([5]).

Why it is hard to be a robot?

However there is a serious gap between the majority of current robots and our expectations about robots. Autonomous arms work in an artificially created, predetermined, static environment where planning can be carried out. Freshly designed mobile robots like lawnmowers wander between electric wires, while owners of automatic vacuum cleaners place deflector walls and remove unneeded objects from the ground to facilitate cleaning.

In contrast to the actual possibilities future robots must work in a continuously changing, fairly complex world, far from this aseptic conditions. It also means they have to create a more viable, human-friendly, or at least more elaborated world model to be successful.

The problem is extremely hard. The results of early artificial intelligence like automatic theorem proving, checker programs, packing bulding blocks in microworlds, in spite of unexpectedly fast and obvious successes, could not yield a major breakthrough in solving less abstract tasks outside the computer with no preprocessed information ([6]).

One aspect of difficulties turns up from the fact that receptors do not deliver direct, human recognizable description of the world, that is they do not generate On(Glass, Table)-like logic expressions, but transport various sets of digital signals to decisioning modules ([7]). The essence of the refinement of our genetically pre-wired perceptional and motional mechanisms does not strike the eye. We can hardly perceive our blind spot, what our vision almost totally compensates. We do not crack up a drinking-glass or let it fall to the ground correcting the holding force continuously and unconsciously. Substantial part of human brain capacity deals with perception and motor control. This kind of intelligent input and output is not given to robots, it has to be created by researchers ([8]).

What humans and robots are common that their perception and action both err ([9]). The applied instruments have finite resolution, they are subject to failure and malfunction, which makes planning more troublesome.

Simulation and reality

Simplification of environment or alteration to the needs of the robot is not an exclusive property of commercial robots. Successful scientific experiments report task solving in university or museum corridors only, robots and their designers rarely dare to go to outdoor spaces. Namely, today's robots work in model worlds instead of real world.

A possible extension of the principle turns up with rise of cheap and fast computers: we can make robots work in virtual model worlds, that is to say we can use special robot simulation softwares. First of all we advance in the direction of simplistic world descriptions which is not suitable for realistic problems, but hopefully, this backtracking makes faster advance possible at a later stage.

Using simulator researchers may build experimental environments according to their own imagination. Complexity, reality, specificity can be gradually increased to a level where virtual robots dare to abandon their cradle and can head to real challenges of the physical world.

One can get rid of lots of tethers by building the robot and the environment virtually, while new problems may appear.

Figure 5: Walking forest machine ([10])

Figure 6: Similar simulated robot ([10])

Backdraws of simulation

Creation of the simulation means some additional computational costs what physical world offers "free of charge": composition of three dimensional world, generation of perceptions, kinematics and dynamics of motion. The simulator has to calculate information appearing on the sensor of the robot, in case of vision including shadows and overlayings. Further task is the preparation of credible trajectory taking into consideration the constraints of the architecture of the robot and paying attention to results of unexpected events like collisions as well.

Figure 7: Research environment ([11])

Figure 8: Industrial environment ([12])

An important threat is that the simulated robot works properly, but because of differences between simulation and reality, the physical version is useless ([9], [13]). To circumvent this problem world properties to be modelled must be determined carefully.

It is worth taking into account that two components are never the same, two motors have different acceleration, two cameras return diverse color maps ([14]). Substantial attribute to be modelled is noisyness and uncertainty of perception and action. Wheels of the robot may slip, the same driving force results various advance speed on carpet or parquet, and even direction of carpet fibres can be significant. A signal emitted by a sonar reflects distinctly from dissimilar surfaces reporting divergent distance information. Because of reflections signal may even reach the receiver of another sonar effecting an evidently incorrect value. Angle of incidence of the signal can be also relevant: in Figure 9 the robot (black circle) equipped with sonars approaches a corner. Measurements are correct if the angle of incidence is perpendicular, but the larger the deviation from this value the bigger the error is as the black line segments show. These facts suggest that real world sensors must be calibrated, and this has to appear in the simulator as well.

Figure 9: Sonar error ([15])

The reverse of the medal is that simulation may induce problems which do not occur in reality ([9]). A simple example is the encounter of two robots in a grid world. In this case the path planner has to solve the stalemate situation. Normally robots do not arrive at the same time at the same position, hence the phenomenon needs no solution.

Advantages of simulation

If there is so much problem with simulation why is it worth using it? A not so scientifical response is that simulation is cheaper. The researcher does not have to pay for a robot, in case of multi-agent simulation, \ for robots. It is not necessary to buy several sensors and other parts and it is avoidable to form a physical but still artificial environment. Moreover a virtual robot cannot be damaged and cannot damage its environment.

A multi-robot experiment also entails technical problems: cables responsible for control and/or energy supply become knotted, or if there is no cable, accumulators of robots must be recharged from time to time ([16]).

Another technically interesting possibility is the parallel execution of the simulation. Researchers may run the experiment with different parameter sets at the same time on several computers, instead of sequential problem solving of one robot. Results will certainly appear faster.

Figure 7: Webots simulator ([17])

Figure 8: Aibos playing football ([17])

A further important factor is that experiments can be assembled faster in simulator than in reality. It is also advantageous in cooperation with real world tests as a fast prototyping of the problem ([18]). Beyond the setup, the execution of the experiment can be more effective. In case of inductive learning algorithms, like artificial neural networks, the real robot may learn only from previously gathered data, since it cannot wander around the environment many thousand times because of its time demand and the deterioration of robot parts. Inside the computer online learning is possible. The problem is more evident with evolutionary algorithms. In this case numerous attempts have to be performed on lots of entities of many generation of a large population to determine the vitality of each individual. For real robots it is only possible with small figures, contrary to a simulator ([13]).

Beside the controller program a simulator may help the development of the robot structure. Type, number, position of the parts, driving mode, and sensor resolution are among the possible variable attributes ([19]).

Another supporting argument of simulation is that side-effects of the divergence from physical world can be diminished. Coevolution of robot and simulator code may result an increasingly real, increasingly efficient robots ([9]).

Simulation or reality? From the above it seems obvious that designers and researchers have to be particularly careful before making a decision. Both methods has strengths and weaknesses, so - as it is usual - the solution can be golden mean. Main part of the robot controller program can be designed in a fairly detailed simulator, and after transferring to reality fine-tuning can be performed. Robots of the future are formed most effectively from cooperation of simulation and reality.

References

1. http://www.kuka.com
2. http://www.parallemic.org/Reviews/Review002.html
3. http://www.robomowcanada.com/
4. http://www.robocleaner.de/english/work1.html
5. http://www.unece.org/press/pr2004/04robots_index.htm
6. S. J. Russell, P. Norvig. Artificial Intelligence: A Modern Approach. Prentice Hall, Englewood Cliffs, NJ, 1995.
7. R. A. Brooks. Intelligence Without Representation, Artificial Intelligence Journal (47), 1991, pp. 139-159.
8. R. A. Brooks. Flesh and Machines, Vintage Books, 2003.
9. R. A. Brooks. Artificial Life and Real Robots, Towards a Practice of Autonomous Systems: European Conference on Artifcial Life, Paris, France, MIT Press, December 1991, pp. 3-10.
10. http://www.plustech.fi/Walking1.html
11. http://innovexpo.itee.uq.edu.au/2003/exhibits/milford/image1.jpg
12. http://www.nosc.mil/robots/images/warehous.jpg
13. N. Jakobi. Minimal Simulations for Evolutionary Robotics. PhD thesis, COGS, 1998. http://citeseer.ist.psu.edu/jakobi98minimal.html
14. S. Nolfi, D. Floreano, O. Miglino, and F. Mondada. How to evolve autonomous robots: Different approaches in evolutionary robotics. In R. Brooks and P. Maes, editors, Artificial Life IV, pages 190--197. MIT Press/Bradford Books, 1994. http://citeseer.ist.psu.edu/nolfi94how.html
15. M. J. Mataric. Interaction and intelligent behavior. Technical Report AI-TR-1495, Massachusetts Institute of Technology, Artificial Intelligence Laboratory, Cambridge, MA, 1994.
16. A. Martinoli. Swarm Intelligence in Autonomous Collective Robotics: From Tools to the Analysis and Synthesis of Distributed Collective Strategies. Ph.D. Thesis Nr. 2069, October, 1999, DI-EPFL, Lausanne, Switzerland.
17. Webots. http://www.cyberbotics.com. Commercial Mobile Robot Simulation Software.
18. Webots: Professional Mobile Robot Simulation, International Journal of Advanced Robotic Systems, Vol. 1, Num. 1, pages 39-42., 2004.
19. Sims, K., "Evolving Virtual Creatures," Computer Graphics (Siggraph '94) Annual Conference Proceedings, July 1994, pp.43-50.