Intelligent robotics - misconceptions, current trends, and opportunities

2.1. Common myths of robotics

There have been many myths, misnomers, and misunderstandings surrounding robots. Some are:

1. Robots have capabilities that equate to or exceed those of humans. This is not true. The intelligence of today’s robots does not exceed even that of a dog! It is inconceivable that robot intelligence can reach that of human intelligence.
2. Robots will steal our spouses, and fight wars, and defeat us. This is very unrealistic and will never happen!
3. Robots will create mass unemployment: people said this concerning the first Industrial Revolution as well. However, in the end, the industrial revolution (and industrial automation) gave the workers more free time (including the five-day workweek), moved them away from hazardous and difficult work environments, improved the general quality of life, and created higher-paying and more challenging employment (for our friends and relative, if not for us).

While many positive things can be said about robotization, the reasons for the underlying misunderstandings are several. They include fantasy, movies, and popular/social media, unrealistic expectations by the robot enthusiasts, slow developments in the field, inadequate expertise and capabilities for the necessary developments, the lack of necessary technologies (mechanical, electronic, and computer science, etc.) to achieve the goals, weak collaboration in the beginning (mainly among the electrical and mechanical engineers, and computer scientists, but this culture has been changed with the establishment of leading robotics institutes); and the lack of adequate “robotic intelligence” for various autonomous operations. See the related video interview of the author^[7].

3.1. Possible directions of advancement

The technology focus may be directed on several aspects to improve the state-of-the-art of Intelligent Robotics. They include autonomy, improved learning and intelligence (for autonomous operation); self-awareness for robots (i.e., knowing the own capabilities of the robot); improved dexterity of handling (e.g., compliant grasping, parallel not sequential, and incoordination; providing the adequate degrees of freedom for manipulation of the handled object); improved robot-human interaction (in particular, working “with” a human rather than working “for” a human); improved speed, stability, robustness, reliability, and safety; improved sensing (particularly, chemical and biological sensing; transparency of remote operations; dynamic sensor networks; intelligent sensor fusion); and significant improvement of the “mechanical” capabilities. Note particularly the capabilities that require “intelligence”, software, mechanical capabilities, and instrumentation.

In this context, a question can be posed whether to direct much effort in developing universal robots that have unlimited capabilities and functions, which will, of course, be very costly and complex as well. In other words, should the focus be directed on the development of very complex and costly multi-purpose robots or use existing single-purpose robots cooperatively? In fact, it is not wise to put much effort into the development of complex and costly robots with numerous features and multi-function capabilities. As a scenario, consider the use of existing low-cost robots that have been developed for just one specific task (e.g., security, human assistance, and guidance, street cleaning). If an emergency occurs (e.g., an explosion), they may be called upon to join (if available) in cooperation, for example, put together simple devices and assist in the situation (e.g., evacuation of the injured).

4.1. Characteristics of intelligence

Before exploring AI itself, let us examine intelligence. No precise definition exists for intelligence. They are the external characteristics and capabilities (that we observe from actions) that enable us to claim whether an entity (for example, a robot) is “intelligent”. Essentially, the outward characteristics define intelligence.

The characteristics of intelligence include sensory perception; pattern recognition; learning (i.e., knowledge acquisition, which is extremely important for intelligence); inference (i.e., making decisions) from incomplete information; inference from qualitative or approximate information (this is commonly used in “qualitative reasoning” as in fuzzy logic or fuzzy reasoning); ability to deal with unfamiliar situations; adaptability to new, yet related situations (through “expectational knowledge”. For example, a human is able to expect the nature of an environment, like a classroom, even when encountering that environment for the first time); inductive reasoning (people must have done this in high school mathematics when proving a mathematical result “by induction”); common sense; display of emotions; inventiveness; and self-awareness (i.e., knowing their own capabilities).

A simplified model for the dynamics of intelligence is shown in Figure 4. The intelligent preprocessors are, in fact, learning modules. They enable one to gain knowledge by “learning” from information and also achieve expertise by further learning through knowledge (including practice). The achieved knowledge and expertise can depreciate for various reasons (including environmental and biological) and also can become outdated. Even though intelligent preprocessing or learning is vital in this model, it is unlikely that machine learning alone will be able to achieve all forms of human intelligence in a robotic device.

Figure 4. A model for the dynamics of intelligence.

4.2. Artificial intelligence

AI uses formal techniques to acquire some characteristics of intelligence. Models of AI are used for this purpose based on one or more of the mentioned characteristics. Such approaches (or models) of AI include Machine Learning, a very popular approach to AI. The conventional models of AI include knowledge-based systems, soft computing (consisting of neural networks - NN, fuzzy systems, and evolutionary computing; see^[9] for instance), and swarm intelligence. A knowledge-based system typically consists of a knowledge base (or a rule base), a database, and an inference engine (the decision-maker). The decisions are made as follows: some data in the database (including what is acquired recently through sensors) is matched with the (context of the) rules in the knowledge base by the inference engine, and the inferences (or actions) are determined accordingly (i.e., rules are fired). Popular “Expert Systems” are based on this model. Of course, the knowledge base will be improved and enhanced continuously through “learning” and experience (so machine learning is used here as well).

Deep learning is a popular approach to machine learning. It incorporates an intelligent and intensive method of learning and sophisticated computing power that is available with such advancements as graphic processing units and tensor processing units to process massive quantities of data efficiently. Deep learning need not be limited to the use of deep NN but is the current trend. Deep NN includes Convolutional NN or convolutional neural network (CNN)^[4,10] (see Figure 5). They have a structure of multiple layers (convolution layers) incorporating the “dynamic” learning ability and ending with a “Softmax” layer, which is the classification layer. First, the NN is trained using “labeled data” (i.e., input data whose proper outcomes are known a priori). Then, after the network is trained properly, unlabeled data (or new data) may be used for actual decision-making. Thus, massive amounts of data, including sensed data (a mixture of labeled and unlabeled data), may be effectively used in a deep NN.

Figure 5. The architecture of a convolutional neural network.

Reinforcement learning relies on rewarding the correct decisions and penalizing the wrong decisions to learn the proper decision strategies. AI agents are capable of providing explanations for their decisions (similar to what Expert Systems provide). In Edge AI, AI algorithms are processed locally on a hardware device. The algorithm uses data created on the device (e.g., data generated by the algorithm) and other data (external data, including those from sensors and through the system interface). Hence, Edge AI functions at the “edge of the system network”. Fuzzy logic attempts to be similar to human decision-making by incorporating “fuzzy” or “qualitative” or “approximate” data, such as those that use qualifiers like fast, small, better, and close. Qualitative or fuzzy reasoning is used in the decision-making (inference) process. Swarm Intelligence behaves like a swarm of animals or insects. They are distributed (not hierarchical) and interact with each other to learn and make decisions. The members in a swarm use very simple rules, yet leading to “intelligent” global behavior, even though an individual member is not quite intelligent, which will improve with time. Evolutionary computing relies on genetic algorithms or genetic computing to realize “optimized” behavior through learning. The basis of this methodology is biological evolution (or survival of the fittest).

Within AI, apart from “learning”, other characteristics of intelligence need to be investigated and incorporated (e.g., decision making with partial, approximate or qualitative information, use of “expectational knowledge”, various approaches of reasoning such as inductive reasoning, ability to deal with unfamiliar situations, common sense, inventiveness, self-awareness, attention representation, and classification). Under machine learning itself, many methods exist, such as CNN, dynamic or recurrent neural networks (RNN), reinforcement learning, support vector machines, and entropy-based approaches. The rationalization of why a particular learning method is chosen (justification) should be a requirement in any application. Comparative evaluations of different methods should be carried out, with a proper reference (basis) for comparison. In this manner, comparative advantages and disadvantages of different methods should be determined, the rationale for choosing a particular approach for the application. When machine learning is applied in a particular situation, domain transformation or domain adaptability needs consideration because the domain of learning is typically not the same as the domain of application^[11].

5.1. Conventional control

The development and application of conventional control have been extended wide effort by many. The relevant techniques include the following.

Feedback control and particularly servo-control of robotic joints had been the main focus in the early developments of robotic control. Here, the robot motions are measured (sensed) and used by the controller in feedback to move the robot in the desired manner. Thus, a robot is “servoed” along a specified motion trajectory through feedback control of the motion error using servo control. Notably, the subject of design and compensation or tuning of proportional-integral-derivative control has received adequate attention.

An image of an object is indeed a valuable source of information about that object. In this context, the imaging device is the sensor, and the image is the sensed data. Depending on the imaging device, an image can be many varieties such as optical, thermal or infrared, X-ray, ultraviolet, acoustic, ultrasound, etc. The image processing methods are rather similar among these imaging devices. For example, the digital camera is a very popular optical imaging device used in various engineering applications such as vision-guided robotics. Such operations as object recognition, pattern recognition and classification, abstraction, and knowledge-based decision making can be carried out using the information extracted through image processing. Visual servoing^[12,13], in particular, has received much attention and is commonly implemented in robotics. Here, the robot motion, including the actual position of the end effector (gripper, hand, tool, etc.) and the relative position of the targeted object, is measured using camera images and compared with the desired or reference values. This difference (error) is used to generate a motion command (feedback control) for the robot so that the end effector would follow the desired trajectory and carry out the robotic task.

A robotic system may have inputs that do not participate in feedback control. These inputs are not compared with feedback (measurement) signals to generate control signals. Some of these inputs might be important variables, while others might be undesirable, such as external disturbances and noise. Generally, the performance of a robotic system can be improved by measuring these (unknown) inputs and using the information to generate the control actions. In feedforward control, unknown “inputs” are measured, and that information, along with desired inputs, is used to generate control signals that can reduce errors due to these unknown inputs or variations in them. The reason for calling this method feedforward control stems from the fact that the associated measurement and control (and compensation) take place in the forward path of the control system. Both feedback and feedforward schemes may be used in the same control system. In some robotic applications, control inputs are computed using the desired outputs and accurate dynamic models for the robots, and the computed inputs are used for control purposes. This is the “inverse model” (or “inverse dynamics”) approach because the input is computed using the output and a model (inverse model). In some literature, this method is also known as feedforward control. To avoid confusion, however, it is appropriate to denote this method as computed-input control.

Since the overall response of a plant (e.g., a robot) depends on its individual modes, it should be possible to control a robot by controlling its modes. This is the basis of modal control. A mode is determined by the corresponding eigenvalue and eigenvector. In view of this, a popular approach of modal control is the pole placement or pole assignment. In this method of controller design, the objective is to select a feedback controller that will make the poles of the closed-loop system take up a set of desired values. This approach uses a “linearized” model of the robot.

As we saw, a robot can be controlled using a feedback control law so as to satisfy some performance requirements. In optimal control, the objective is to optimize a suitable objective function (e.g., maximize a performance index or minimize a cost function) by using an appropriate feedback control law^[14]. A particularly favorite performance index is the infinite-time quadratic integral of the state variables and input variables, and popular control law is linear constant-gain feedback of the system states. The associated controller is known as the linear quadratic regulator (LQR). Linear Quadratic Gaussian (LQG) Control is an optimal control technique that is intended for a linear system with random input disturbances and output (measurement) noise. An LQR controller together with a Kalman filter is used in this approach.

For servo control to be effective, nonlinearities and dynamic coupling of the robot must be compensated faster than the control bandwidth at the servo level. One way of accomplishing this is by implementing a linearizing and decoupling controller inside the servo loops. This technique is termed feedback linearization technique.

An adaptive control system is a feedback control system in which the values of some or all of the controller parameters are modified (adapted) during the system operation (in real-time) on the basis of some performance measure when the response (output) requirements are not satisfied. Many criteria can be employed for modifying the parameter values of a controller. Self-tuning control falls into the same category. Model identification or estimation may be required for adaptive control, which may be considered to be a preliminary step of “learning”. A neural network may be used for this purpose. In a learning system, control decisions are made using the cumulative experience and knowledge gained over a period of time. Furthermore, the definition of learning implies that a learning controller will “remember” and improve its performance with time. This is an evolutionary process that is true for intelligent controllers but not generally for adaptive controllers. In model-referenced adaptive control, the same reference input that is applied to the physical system is applied to a reference model as well. The difference between the response of the physical system and the output from the reference model is the error. The ideal objective is to make this error zero at all times. Then the system will perform just like the reference model. The error signal is used by the adaptation mechanism to determine the necessary modifications to the values of the controller parameters in order to achieve this objective.

Sliding mode control^[15], variable structure control, and suction control fall within the same class of control techniques and are somewhat synonymous. The control law in this class is generally a switching controller. A variety of switching criteria may be employed. Sliding mode control may be treated as an adaptive control technique. Because the switching surface is not fixed, its variability is somewhat analogous to an adaptation criterion. Specifically, the error of the plant response is zero when the control falls on the sliding surface.

H∞ (H-infinity) control is an optimal control approach, which is different from the LQG method. This frequency-domain technique assumes a linear plant with constant parameters, which may be modeled by a transfer function (matrix in the general case). The underlying design problem is to select a suitable controller that will result in the required performance of the system. In other words, the closed-loop transfer matrix must be properly “shaped” through an appropriate choice of the controller. Specifically, the controller that minimizes the “H∞ norm” of the closed-loop transfer matrix, which is the maximum value of the largest singular value of this matrix, is used.

For complex multi-robot systems having various and stringent operating requirements, distributed and networked control is appropriate. It may consist of many programmable logic controllers and a supervisory controller, which will supervise, manage, coordinate and control the overall system. In hierarchical control, the distribution of control is provided both geographically and functionally. The management decisions, supervisory control, and coordination between robots may be provided by the supervisory controller, which is at the highest level of the hierarchy. The next lower level may generate control settings (or reference inputs) for each control region (subsystem). Finally, setpoints and reference signals are inputs to the direct controllers of the robots. In master-slave distributed control, only downloading of information is available.

5.2. Intelligent control

In intelligent control, an “intelligent” method of decision-making is used to make the control decision (i.e., to generate the control action). Soft computing, consisting of neural networks, fuzzy systems, evolutionary computing, and even probabilistic methods, has been popularly used in intelligent control. The topic of soft computing has already been addressed under the general theme of the present paper and is not repeated here. However, it is adequate to mention that, since learning control is used in robotic control, any approach of machine learning such as deep learning and deep neural networks, as discussed earlier in the paper, maybe incorporated into intelligent control of robots.

6.1. Opportunities for developing countries

Here, opportunities exist in all the areas that were mentioned before. However, developing countries should not blindly decide on the considered robotics activities just for the sake of being involved in Robotics or AI. It is important to explore and determine what is in the “black box”. Otherwise, one can be dissuaded through fear-mongering or make wrong choices for robotic activities. One must first question whether Robotics is needed for a specific local application. Then they must explore which robotic approaches are relevant for the considered task. Very importantly, they must examine what is in the existing Black Box before implementing it.

Developing countries should primarily concentrate on “robot development”, not their application for the automation of local industries. They will be able to market these robots to other countries. Since the developing countries typically have an excessive and smart labor force, using robots for such applications as agriculture and industrial automation is not generally suitable in those countries. Nevertheless, they may consider the development of simple and low-cost robots for local use (e.g., for service and household applications). They may focus on the development of advanced software, in particular, to incorporate other forms of intelligence into robots and efficient software, and the use of advanced platforms like Flexible Cloud, Real-time Internet of Things, and Edge AI. Software development can be carried out without much capital investment, as it is done in India particularly because these countries normally have an educated and vast group of professionals. Also, they should focus on advancing the “mechanical capabilities” of robots, which are essential but may not necessarily be for the local market. As well, they should consider the needs that result from a particular situation (e.g., Covid-19). Very importantly, they should develop their own guidelines and regulations for robotic ethics and safety, which can be done by modifying the existing guidelines and regulations in the highly developed jurisdictions.