Perception in robotics

Citation

Sanders, D. (1999), "Perception in robotics", Industrial Robot, Vol. 26 No. 2. https://doi.org/10.1108/ir.1999.04926baa.002

Perception in robotics

Perception in robotics

The author

David Sanders is Principle Lecturer in Mechanical and Manufacturing Engineering, University of Portsmouth, Portsmouth, UK.

Keywords Machine vision, Robots, Sensors

Robotics has come a long way during my lifetime. Throughout all the changes and advances I felt that I understood some distinct differences between robots and human beings. The differences were obvious when Joe Engelberger set the first Unimate robot to work at General Motors. Such robots were clearly our slaves. We are now entering a new and exciting age where I am not so certain of these distinctions.

To advance further in robotics we need to improve sensing in order to make robots more user-friendly and capable... and it is happening. Almost every volume of every science or engineering journal contains articles about new sensors or new ways of using sensors. At the same time, the information from sensors is being fed back into systems at a higher and higher level (see Figure 1). It is these improvements in sensing which are blurring my understanding of what it means to be a robot and what it means to be a human.

I will try to explain this.

Sensing involves transducing and then processing data or information. Until recently, a robot tended to sense things directly while we as human beings tended to sense things through our natural human senses: vision, touch, taste, hearing, or smell. Robot sensing had the advantage that it could be interfaced directly while natural human sensing had the advantage of being more perceptual. For example, consider a temperature sensor. A robot may have a temperature sensor connected directly into the robot controller but a human being must detect the output of the sensor through a natural human sense and then process the data. For example, the human might see a thermometer or hear a fire alarm (Plate 1).

This may change! The performance of silicon microchip processors and memory is still doubling every 18 months. High-end processors can run at 750 MHz with four million transistors per chip. By 2003, they are expected to be running at 1500 MHz with 18 million transistors per chip. New processors may have between 5 and 20 clock instructions per clock cycle, that is an equivalent of 10 billion instructions per second per chip!

As computer processing increases in power so robots may improve their perception and learn to use new tools and sensors. Human implants may also become more common and humans may communicate directly with external sensors or machines through other means.

Figure 1 Feedback from sensors is being fed back into systems at a higher and higher level

Plate 1 A human being must process external inputs from sensors and provide control using natural human senses, for example: vision, touch and hearing

For both humans and robots, the most important sensors are often visual. When vision is not available then the lack of information is often overcome by placing things in fixed locations and orientations. This can be expensive and difficult. A lack of sight restricts the options for the machine or human being and other senses have difficulty in making up for the deficiency. Vision and image processing make life easier and allow for more complexity and flexibility.

In robotics, the visual sensor is often a TV camera. Much research effort around the world has turned to computer vision, mostly driven by the end-user market. It is here in image processing research that some of the greatest advances are being made. But robot vision sensing still needs to advance in two main directions: performance needs to be improved and information from the sensors needs to be more easily mixed with information from other sensors, such as ranging or proximity sensors. Chip, lens, speed of adjustment and fibre optic technology can still be improved and in pre-processing, signal to noise ratios could be improved in a variety of situations. But the real advances will come with new methods of integrating data and information.

In a previous Viewpoint, Dr Kerstin Dautenhahn of Reading University identified language as being key in allowing the communication of ideas, plans and identity, as well as data and information. The search for a universal language is related to the search for universal software architectures, which may be needed to achieve the sensor and control integration. The research in image processing may provide the steps towards this sensor integration.

Computer scientists and software engineers are making rapid progress with research in object-oriented languages and parallel processing. Although the code produced by these languages may sometimes be less efficient than hard coding, it is generally more flexible. The idea of parallel inheritance from existing classes to new classes may be powerful for the control and understanding of robots and for the way in which sensors are treated and integrated in systems. In any case, the new way of thinking involved in this will lead to new methods and ideas for the simple integration and representation of information from sensor data.

If we can achieve this integration then many of the problems we think of today as high level problems may just become low-level nuisances. For example, dual robot co-ordination may become simple. Work may move to a higher level to concentrate on the application or the task. A multitude of robots may be treated as one entity. In image processing, the recognition and positioning of different objects is becoming easier, so that systems can concentrate on what they are going to do with objects rather than what they are and where they are. That is, to concentrate on the task rather than the component parts.

All this depends not only on improving sensing to make robots more user-friendly and capable, but also on the ability of the robot to understand tasks in real time, so that they know what is happening as it happens. This includes advances in modelling tasks and the environment and the integration of the two in order to generate new algorithms. These algorithms can be used for reasoning about relationships between objects and tasks and for inferring the results of possible actions. If sensor inputs and physical obstacles and tasks can be represented as software-objects then a robot may know for example that a collision will occur if it moves left. It may not have to calculate it at all as the task and the obstacles are all part of the whole.

Once we get there then perhaps a whole new host of problems will face us. We will need to guide and communicate rather than command the machinery.

As robots misunderstand, make mistakes and apply more creative solutions to problems then how will we tell the difference between them and us?