11.1 Technical Limitations

Despite all of the development in terms of base technologies, only now are we beginning to devise how sensing, state inference, and actuation can be combined together in HiTLCPSs, as evidenced by the research projects described in Sections 4.1 and 4.2. In general, most of these projects still assume architectures within environments that are well known and static. We believe that future IoT environments will be mobile, dynamic, and reactive, where humans and technology will have to react in real time to stimuli from the environment, in order to guide their actions [186]. To this end, WSNs allow for the monitoring of environmental conditions, helping IoT devices, robots, and humans to react much more effectively to changes.

Additionally, most current scenarios do not fully consider humans, their behaviour, and their psychological state as integral parts of the system. Humans are still mostly seen as an external final user, and rarely directly interfere in the control loop of working tasks. As far as we know, there is no significant work that fully utilizes the potential of the human element to support the control system itself. In all previous projects there is a very well-defined border between humans and the system, instead of a tightly coupled integration. As we saw in Chapter 3, humans can play various roles within HiTLCPSs, ranging from actuators co-helped by robotic elements and acting on information collected by the sensor networks to intermediate nodes in multi-hop communication processes. They can also become an element of environmental monitoring (through the sensors carried by them, e.g. on smartphones or smartshirts).

The pieces of work presented in [74], [124], and [16] are quite good demonstrations of the potential of HiTLCPSs. Still, we have some reservations about the feasibility of the presented approaches for widespread deployment. The use of vision-based systems is very prone to noise and limitations in image processing, only working for very controlled environments (e.g. recognizable objects limited to those programmed into the system). Brain/computer interfaces based on EEG signals are not practical, since electrodes are usually very cumbersome to wear and thus not suitable for day-to-day HiTL applications. This leads us to believe that near-future HiTLCPSs will most likely be based on more pervasive and mobile technology. In particular, the smartphone is a ubiquitous sensing platform that is already used by millions of people around the globe, every day. Theis device gives us the sensing power and computational capabilities that might be key for the first generation of massive HiTL deployments coming in the next few years. Still, there are few actual applications of smartphones and HiTLCPSs. While [118] did use HiTL concepts to limit the current mobile data demand, the actuation aspect was limited to suggestions and incentives on a smartphone's graphical user interface, and aspects such as robotics and direct actuation were not considered.

Let us attempt to condense all of these technical limitations and challenges in to a single model, shown in Figure 11.1. This model presents the various processes associated with HiTL control. A human is integrated into a CPS through “human-in-the-loop intelligence”, responsible for receiving input from the human sensors and also for influencing the system's control loop depending on the inferred context. This intelligence's specific implementation should follow the general principles and requirements introduced in Chapter 10, to guarantee reliable and secure human-context monitoring. In particular, we consider the issues of privacy and reliability as two of the most important requirements, largely responsible for the current lack of HiTLCPSs in real scenarios.

In a first step, determining a human's state requires the acquisition of data, through sensors (carried by the human or present in the environment) and/or from information present in social networks. This information can relate to several aspects of physical reality, such as the human's thought patterns through EEG, who their friends are, their heart rate, movement through accelerometers, positioning through GPS, and facial expressions through video-cameras. Assessing physical reality through sensory data is the cornerstone of HiTL control, since every other aspect of the system is related to the raw data acquired from the sensors.

History, or memory, is another important aspect that is closely relate to the acquisition of data. In fact, research has shown how previous human states may provide important insights for inference mechanisms [187]. This historical data can also be used by delay-tolerant mechanisms in non-critical applications, setting a meaningful state whenever the real-time connection to sensory data is interrupted.

Perhaps one of the most critical aspects of HiTLCPSs is the reliable inference of human state. State inference mechanisms need to adapt to the current context as well as the human's preferences and historical behavior, integrating this information into the control-loop as feedback to determine the actions of the HiTLCPS. This is extremely difficult and implies a need for reliable and secure mechanisms for modeling, detecting, and possibly predicting human nature, as discussed in Section 4.1.2.

There are two types of actuation in HiTL controls. A system actuation is based on the system's current status and the inference of human state. For example, an HiTL-enabled HVAC system should only adapt the room temperature in the presence of humans. Human actuation relates to the actions of humans within the HiTL system, since they can themselves actuate whenever necessary. Motivation is a crucial aspect of this type of actuation and one of the most important research challenges. Future HiTLCPSs need to provide the necessary motivation and benefits for humans to act in a way that benefits the overall system and refrain from adopting greedy or prejudicial attitudes.

Finally, noise reminds us how real-world environments are far from idealized academic-controlled testbeds. For example, HiTLCPSs based on speech and video-captured gestures have to deal with challenges such as ambient noise, moving background clutter, and object segmentation. The acquisition of human vital signs is also prone to problems in terms of signal-to-noise ratios, where many signal frequencies result from internal physiological functions that have nothing to do with what needs to be acquired.

Another source of noise relates to human variability. The human species has a high genetic variance and thrives in many different environments with highly disparate cultural backgrounds, which results in many possible phenotypes. Age, physical disability, and interperson variability also need to be taken into account. While current research in HiTL state inference can reach high levels of accuracy, as discussed in Section 4.1.2, these results are mostly limited in terms of number of human activities, psychological states, and audience. On the other hand, future HiTLCPSs will most likely address a highly heterogeneous target audience. This personalization of existing state-inference models should follow a ubiquitous approach, and not overly depend on manually providing training examples or on the collaborative labeling by the system user. To promote usability, it should be a transparent process that happens naturally, as the user lives his/her daily life.

The identification of new human states that were not predicted at the time of deployment may also be important. However, this brings yet another realm of unresolved challenges. It is necessary to scale the learning of new states, avoid redundant labeling, perform training in a lightweight fashion, ensure security and privacy, and take advantage of collaboration between users while avoiding overlapping efforts.

All of these are important challenges for HiTLCPSs that have yet to be properly addressed by research in the field.

11.2 Ethical limitations

As discussed in Section 4.1, much of the necessary technology for supporting HiTLCPSs is already in place. But then, why are current IoT and CPSs still unable to integrate the human element into the control loop? As previously discussed, we believe that reliability is one of the major factors that influence the current lack of real-world deployments. Reliable and consistent inference of a human's state are essential for the adoption of HiTLCPS in real industrial, medical, or social scenarios. The inability to do so can have severe consequences on the effectiveness of the entire system. Reliable networking is also crucial for HiTLCPS, since these systems are often distributed and need to share information between many devices.

There is, however, another important factor that needs to be taken into consideration: the introduction of radically new technologies is usually accompanied by a considerable dose of skepticism. Thus, reliability is only relevant if the market accepts the underlying technology. This is crucial, since this new paradigm of human-centric technologies has already been previously met with some reservation. As evidenced by Section 2.2, current attempts at creating social-networking HiTLCPSs show that users place a high importance on their privacy and on the security of their personal information. In fact, these privacy concerns have been present since the beginning of social networking. Facebook, for example, has been the target of criticism since its beginnings because of its reliance on the user's willingness to share information as the key point of its business. In fact, according to an AP-CNBC pool [188] with a sample of 1004 people, 59% of Facebook users have little to no trust in Facebook to keep their information private. This apparent lack of trust reflects just how closely people follow intrusive practices, further exemplifying how privacy concerns are one of the biggest obstacles to the growth of social networking and, by extension, to HiTLCPSs.

Concrete examples of how such reservations may affect the introduction smartphone-based HiTLCPSs can be seen in the cases of Highlight and SceneTap, previously presented in Section 4.2.3. Concerning Highlight, Baig [189] argued that such encounters are sometimes best “left to fate” and that the application “may tell others too much about you”, but still praised its functionality and novel features.

As for SceneTap, skeptics advocate privacy concerns and have raised questions around the facial detection technologies used to collect information, since they are employed without people's consent. The application met a troubling launch in May 2012, where it was supposed to be supported by 25 San Francisco bars of which ten dropped out after angry calls and an editorial that called the service “creepy”. The app has also been criticized for its gender filtering options, letting people find bars with a larger proportion of men or women in a certain age range [190]. In addition to the ethical concerns, SceneTap was also plagued with technical limitations, as pointed out by Anderson [191]: the app apparently had accuracy problems with its facial recognition software, with several bars showing high attendance levels when in fact they were “as dead as can be”. She also claimed that the software failed to register her presence when she attempted to enter a bar very slowly.

Putting skepticism aside, it is difficult to deny that the idea of someone else monitoring our every step and activity is very disturbing. However, it is also true that this problem does not reside entirely with the existence of HiTLCPS frameworks. For example, Sauvik et al. [192] discussed the possibility of current smartphones posing a security threat to the user, claiming that accelerometers and other sensors within the device could be used without the user's consent. They have also shown how activity recognition algorithms can be used to obtain sensitive information about users without their knowledge, by having them identify pre-defined general activities or even make the user's phone learn to identify new ones. Hence, the existence of smartphone-based HiTLCPSs does not impede this type of privacy invasion, although it might make it easier to accomplish.

Still, it would have been, perhaps, unthinkable in a pre-social-networking era that people would enjoy publishing their personal information in a public database for their peers to see and comment on. Yet, little by little, we have reached the stage where huge social networks and photo sharing are the norm. Despite all the past and ongoing privacy concerns and surrounding criticism, both the number of users and their engagement in mobile social networks continue to increase [73].

Nevertheless, security and privacy are two other critical requirements, in addition to reliability, for HiTLCPSs. Industrial processes, medical data, and sensitive personal information need to be protected from unauthorized exploitation. As already discussed, protecting confidential information is often not only a business requirement but, in many cases, also an ethical and legal requirement.

Another important ethical consideration relates to the use of robotics in HiTLCPSs. As mentioned in Section 4.1.3, robotics is growing at a progressively faster pace and there are some who believe its role in future HiTLCPSs may not be completely optimistic. For example, while robotics enables automation, this may in turn result in human unemployment. In fact, futuristic journalist Kevin Kelly predicts that a wave of automation centered on artificial cognition, cheap sensors, machine learning, and distributed intelligence will likely result in 70% of today's occupations being replaced by automation before the end of this century. Starting with assembly line and warehouse work, agriculture picking, cleaning, “it doesn't matter if you are a doctor, lawyer, architect, reporter, or even programmer: The robot takeover will be epic” [107].

Brynjolfsson and McAffee provide an interesting insight into this matter, arguing that despite the improvement of technology in areas that used to be typically human-oriented, such as pattern recognition, people will still have vital roles to play [67]. As an example, they refer to Garry Kasparov's experience in “freestyle” chess tournaments, where teams combining average-skilled humans and machines dominated both strong computers and human grandmasters [193]. As pointed out in Diego Rasskin-Gutman's book, Chess Metaphors, what computers are good at is where humans are weak, and vice versa [194]. This is evidence of the importance of human–machine collaboration in the years to come, the cornerstone of HiTLCPSs. Brynjolfsson and McAffee continue their discussion on these “uniquely human” abilities that will remain essential, even in the face of the continued automation of routine tasks by technological advancement. Despite their impressive calculation capabilities, there has yet to exist a machine that is capable of human creativity and intuition. The ability to create and innovate through new and meaningful ideas preoccupies the pinnacle of AI (AI) research, and is the one task that humans still excel in comparison to machines. Additionally, evolution has shaped humans into highly responsive beings that can quickly adapt to new situations, while current machines simply cannot react outside of the frame of their programming. As evidenced by Brynjolfsson and McAffee, “[The supercomputer] Watson is an amazing Jeopardy! player, but would be defeated by a child at Wheel of Fortune, The Price is Right, or any other TV game show unless it was substantially reprogrammed by its human creators” [67]. Thus, human–machine collaboration will most likely become increasingly critical in the next few decades, at least until machines evolve to a point where they reach (or surpass) “human-like” intelligence. As memorization skills become increasingly redundant owing to the assistance of modern search engines, it is this human ability to quickly combine information from different sources and to react to new situations that will remain essential in future HiTLCPSs.

Precursors of this human–machine interaction are already among us. Baxter, a workbot from Rethink Robotics, is an early example of a new class of industrial robots created to work alongside humans [195]. Baxter has several characteristics that make it more “human-aware” than most of its ancestors. It is capable of showing where it is looking by shifting drawn eyes on its “head”. It is also capable of perceiving humans and avoid injuring them, using force-feedback mechanisms that tell it it is colliding with a person or another bot. This “human-like” body language is an innovation that allows humans to understand and predict the robot's intentions, which may in turn reduce the previous mistrust placed in robotic companions [108 109]. Equally important is Baxter's capability of learning through imitation: to train it, one simply grabs its arms and guides them through the correct motions and sequence. This mode of operation is remarkably different from traditional industrial robotics, which requires highly educated personnel to program even the simplest tasks. Considering all of these tendencies, it is very likely that, in the future, people will be paid “based on how well they work with robots” [107].

Another good example of how robotics and HiTL are becoming intimately related is Pepper, a humanoid robot designed by Aldebaran Robotics and SoftBank Mobile who is capable of reading human emotions [196]. Unlike the Baxter workbot, Pepper is an emotional robot, not a functional robot. It was designed with the purpose of making people happy, making them grow, enhancing their life, and facilitating their relationships. To do so, it is capable of communicating through “voice, touch and emotions”, maintaining conversations, and “having fun”. For example, Pepper is capable of understanding laughter and associate it with “good mood”. It does this through knowledge of “universal emotions” (joy, surprise, anger, doubt, and sadness) and its ability to analyze facial expression, body language, and language. Its also adapts its behavior according to the situation, and employs machine learning to better understand humans, being capable of learning its user's tastes.

It is very likely that Baxter and Pepper are just the first signs of a new technological revolution that is quickly approaching. On one hand, expecting AI to evolve until it becomes “humanlike” is “the same flawed logic as demanding that artificial flying be birdlike, with flapping wings” [67]. It has already been proven that tremendously complex programs, despite being based on simple instructions, are already able to outperform human thinking. On the other hand, innovative, creative, and imaginative machines are yet to be seen and it is unlikely that humans will be replaced in this department in the decades to come… or is it?

Developments in AI research are continuously casting shadows of doubt on the prowess of the human mind. One of the last bastions of the human mind against AI evolution was the ancient game of Go: originated in ancient China, more than 2500 years ago, Go remains the oldest board game still played today. Unlike chess, which has long been beaten by machines, Go had, until recently, been considered a tremendous challenge for AI. As put by mathematician I. J. Good in 1965 [197]:

In order to programme a computer to play a reasonable game of Go […] it is necessary to formalise the principles of good strategy, or to design a learning programme. The principles are more qualitative and mysterious than in chess, and depend more on judgment. So I think it will be even more difficult to programme a computer to play a reasonable game of Go than of chess.

In fact, for a very long time, most computer Go programs were considered worse than an average player with just a few years, experience; Go is a game that may take an entire lifetime to master. However, that all changed very recently.

In January 2016, Google published a paper on how they managed to build an AI, named AlphaGo, that won a match against a professional Go player: the European champion Fan Hui [198]. Hui, ranked 2-dan, lost a five-game match at the Google DeepMind office in London in October, without handicaps. In March 2016, AlphaGo and South Korea's Lee Sedol, considered one of the highest-ranking Go players of the last decade, played an historic five-game match. AlphaGo won the match (4–1), making it the first time a computer Go program defeated a world-class human player on even terms [199]. Mankind may have just lost its last bastion against computer intelligence, at least as far as board games are concerned.

However, what about scenarios where machine intelligence can have a more dire impact on human lives? This is the case for autonomous vehicles (AVs), which are about to become a tremendously important type of CPS. Projects such as Google's Self-Driving Car [200] are racing ahead, trying to create fully automatic cars that simply require no human driving whatsoever. It is rather interesting to consider the possibility of such vehicles adapting their behavior or their interior environment to their occupants' desires. Should a self-driving car attempt to go faster if its owner is in a hurry? What about slow leisure drives across the country? Could a driverless car be capable of adapting its driving behavior and route in order for its occupants to enjoy the scenery?

One of the most prominent reasons in favor of automated vehicles is the possible reduction of deaths from traffic accidents. However, in these HiTL scenarios, the consequence of failure certainly has a greater impact than losing a freestyle chess match or mis-detecting a human emotion in a smartphone app. What if, in its attempt to please its owner by changing its driving condition, the vehicle ends up being forced to choose between two evils? In a situation where an accident is inevitable, should the vehicle run over pedestrians or sacrifice its passengers? This dilemma is interestingly discussed by Bonnefon et al. in their report The Social Dilemma of Autonomous Vehicles [201]. The authors pointed out how the potential consumers may be more willing to ride in AVs that protect their passengers at all costs.

How can this selfishness factor be input into the design of decision-making algorithms for AVs? Will car manufacturers favor AI that values the desires and safety of their passengers over other individuals? According to Bonnefon et al., manufacturers and regulators will need to accomplish three potentially incompatible objectives: being consistent, not causing public outrage, and not discouraging buyers [201]. These are difficult ethical decisions that will have a profound impact on the adoption of this and other types of HiTLCPS technology that can influence human integrity.

From all of this, we can safely say that while intelligent HiTLCPSs will most likely “think” very differently from us, at the same time they will further integrate humans and their intuition into their own control-loop tasks. Without a doubt, for better or for worse, HiTLCPSs are here to stay and will become increasingly more prominent and ubiquitous in our daily lives. In the face of this, it is now high time that we, humans, take decisions and act, and do not limit ourselves to observing the long-term consequences of such systems and how they will transform our world and the way we live.