This section of the blog is dedicated to the heritage of Rodney Brooks to embodied cognition. Struggling to apply the traditional AI to intelligent robotics, he developed new approaches and new concepts. Each new concept is developed in a distinct subsection.
3.1. Behavioral Robotics
Rodney Brooks developed the model of layered architecture in his early papers published during the ’80s (Figure 3.1.1). The layered architecture is divided into behaviours and not in functional units. This approach makes it easier for example to add a new functionality. One layer is implemented and then tested, and only when this layer works the way it is supposed to, another one can be added. The global behaviour of the robot is more likely to meet the specification than in the architecture based on functional units. They can be tested incrementally, at the end of the implementation of each layer. This is useful to have a behaviour conform to the one in the specification of the robot.
Figure 3.1.1: The functional modules decomposition and the layered control system ([11])
These tests demonstrate the robustness of the robot. Since robots are usually in contact with people, they need to have a behaviour that is very precise so that it does not harm anyone. In the cases in which the robot is not in contact with people, its behaviour needs to be precise so that he does not harm himself and get broken. If the tests are successful, the robot will not have many bugs and behave well in most situations. Such tests are for instance critical for the robots that assist nurses in surgical operations.
The Kaspar robot ([28]) and the Roomba robot (Figure 3.1.2) are both embodied systems that interact with children and eventually animals for the second. They have been constructed using the subsumption architecture, which made them robust to their environment. This is necessary for the Roomba robot which is a vacuum cleaner and that is confronted with a changing environment like a house in which people can come in the way. It is not supposed to bump in them. It also needs to be capable of going to his charging base when needed. Kaspar it is a humanoid robot able to interact with children, especially autistic one to help them for example learn social rules. It also needs to be robust to his environment.
Figure 3.1.2: Roomba robot ([76])
Robots like these are more and more present in our lives because they are commercialized much more than before. They are safer and more efficient.
They still are robots that are not commercialized in supermarkets, but that are used in companies where help is needed. They have precise behaviour so they can help efficiently the workers. Other robots of this type can also be found in hospital, for example for surgeries.
Actually, those intelligent robots cannot be conceived by dividing their conception into subproblems. It means that one shouldn’t design them step by step until conceiving an intelligent system.
3.2. Intelligence without Representation
Figure 3.2.1: Rodney Brooks & Cog the android robot [63]
“Look at an insect, it can fly around and navigate with just a hundred thousand neurons. It can’t be doing this very complex symbolic mathematical computations. There must be something different going on.” said Rodney Brooks in a 1997 interview [15], six years after the publication of “intelligence without representation”, a paper in which he discusses the philosophical underpinnings of research in AI, and exposes his views on how intelligent systems should be built.
Since their inception, cognitive science and artificial intelligence have been dominated by computationalism and functionalist theories, which suggest that the mind is an information processing system. However, these theories were reconsidered because of arguments such as The Chinese Room, stating that AI programs were not intelligent in the strong sense (equal to human intelligence). Through the 80s and the 90s, an alternative approach endorsed by Brooks emerged, embracing the principles of a situated and embodied cognition. This approach represented a more radical conception of cognition with AI arising out of biology and the study of animal behavior, and led to the writing of “Intelligence without representation”.
First, it is interesting to know more about the history of this paper. Its original title is “Achieving Artificial Intelligence through Building Robots” [11]. It was written in 1986, shortly after “A Robust Layered Control System for a Mobile Robot” [10] discussed in the previous section”. It was not well received by its peers and went through a series of rejections. However, this paper caused quite a stir in the scientific community, and its fame eventually spread. In the end, it was published a few years later by the Artificial Intelligence Journal, the mainstream journal of the field [20], and totals 7238 citations at the moment (May 2021).
Brooks thesis is that traditional AI research addressed the wrong type of intelligent behavior. At that time, traditional AI decomposed intelligence into specialized subproblems such as knowledge representation, natural language processing, or computer vision. However, he believed that human intelligence is too complex and little understood to be broken down into the right sub pieces. Efforts should be spent on modeling simpler intelligent systems instead of trying to glue all the different sub pieces together hoping that they would fall into one place.
Brooks used the following analogy to illustrate this idea: imagine that there are researchers in the 1890s working on artificial flights, who are transported in the 1980s using time travel on an airplane for the duration of a flight, so that they can experience the fact that artificial flight is possible. On their return, the researchers work on replicating the same aircraft they have seen but without necessarily trying to understand the laws of aerodynamics. Brooks believed that the AI researchers are similar to these artificial flight researchers as they aimed at reproducing human intelligence without the necessary foundations or knowing how our intelligence is developed from the most basic levels.
Another topic that bothered Brooks very much was the notion of intelligence representation. Methods in traditional AI research were based on the symbol system hypothesis, which assumes that many aspects of intelligence can be achieved by the manipulations of symbols. Brooks disagrees with this approach; he argued that robot should not represent the world via an internal set of symbols and then act on this model, and claims that “Representation is the wrong unit of abstraction in building the bulkiest parts of intelligent systems”.
He believed that intelligence is demonstrated through interactions with the real world, which means that proper perception-to-action setups can be used to directly interact with the world as opposed to modelling it. He wrote that it is better to use the world “as its own model”. By getting rid of any central representation, Intelligence without representation introduces a different paradigm to modelling intelligence, hence the disruptive nature of this paper.
The two cornerstone ideas of this behavior-based approach are situatedness (physical world directly influences the behavior of the robot) and embodiment.
Figure 3.2.2: Visual sketch of embodied cognition: the body, the mind and the environment all work together to influence each other [104].
In biology, the concept of embodiment can be divided in two different types: Loebian (behaviorist embodiment) and Uexkullian (phenomenal embodiment) [81].
Loebian embodiment follows the behaviorist line that Brooks supports. In this case, cognition is embodied in the control architecture of a sensing and acting machine, where the behavior is grounded in the interaction between agent and environment.
On the other hand, Uexkullian embodiment appears as a quite different type of embodiment in which each animal species perceives the world differently. It roots on bringing together an organism’s perceptual and motor worlds to form a closed unit – the Umwelt (subjective or phenomenal world). Therefore, this type suggests that subjectivity plays a role, and that similar organisms can behave differently even though they share the same environment.
Prior to “Intelligence without representation”, the goal of artificial intelligence was to simulate human intelligence in a computational system. Given the lack of complete understanding of how intelligence and our mind work, Brooks introduced a new approach, from which the field of behavior-based robotics emerged. It allows the robots to adopt a complex behaviour without programming everything.
3.3. Interaction
The term embodiment has not always achieved a consensus in the scientific community, since some argue that it is simply a physical existence of a body associated with a controller, and others support that the interaction of the embodied robot with the surrounding environment and its continuous adaptation to obstacles must be achieved more than ever before to improve robustness and the ability to cope with real-world problems (Figure 3.3.1).
Figure 3.3.1: Robust Embodied Intelligence [97]
The interaction between the intelligent system and its environment is key to embodied intelligence. There are two different ways to reach the interaction between the system, or more specifically, the robot’s control architecture, and the environment to generate behaviour. The traditional way is based on a top-down approach where the processing is decomposed into a chain of information processing modules proceeding from sensing to action, resulting in high-level reasoning skills but lacks real world robustness ([13]). On the other hand, there is the bottom-up approach where the decomposition is made in terms of behaviour-generating modules, in other words, layers of competence are built, and they can be working on individual goals concurrently. Due to this layered architecture already explained in detail in the section 3.2, multiple sensors can be added to the system, increasing the robustness of the robot, which is also increased by the possibility of the lower levels produce results even if the higher ones can’t do it in real time ([10]).
With the increasing attention given to this topic, it has become imperative to better understand the concept of embodiment. Initially it was believed that the body was only a vehicle for the mind and that essential attributes of humans were exclusively attributes of the spirit. In contrast, more recent theories argue that embodiment and intelligence are inextricably linked. Each natural system experiences different physical stimulus during its lifetime and this different stimulus give them distinct learnings from it to adapt to a complex nondeterministic environment which is reality. Our body affects how we develop new skills. For example, the shape of our feet and hands allows us to walk on two legs and use a cell phone at the same time. A task that a dog, for example, could not perform.
Starting from the idea that the body is an essential part of learning and adaptability of any intelligent organism, it remains to be understood how an animal can be distinguished from an embodied intelligent system. The answer rests on a distinction between two terms: autopoiesis and allopoiesis. The first one, autopoiesis, refers to systems that can adapt to their environment at both macro and micro levels. The second one, allopoiesis, refers to systems that can only adapt at macro level. Since an autopoiesis system is more autonomous and more capable to respond to changes in the surrounding environment, the goal became to bring robots as close as possible to this type of approach to make them more and more robust and autonomous. This is sometimes called In-World embodiment, in contrast to to the ON-World embodiment, which corresponds to an allopoietic interpretation. IN-World robots interact with their environment in a dynamic and adaptive way, allowing real-time responses, which will be explained in more detail in the next section. The latter approach does not require the robot to have all the “answers” inside it, allowing it to learn from the environment and adapt to unpredictable circumstances. Autopoiesis is therefore closely connected with the concept of an autonomous system.
The idea that a brain develops inside a body that can interact with the world gave birth to embodied intelligence and is what makes a robot different from a computer that cannot explore and interact with the physical world. The robot iCub, a humanoid robot the size of a small child is a good example of an effort to develop a more robust robot that is more capable of adapting from a body. The goal in building this robot was to understand to what extent it would be possible to create a robot that would learn over time just like a human being. The robot iCub has a child-like body so that it can learn similar skills and make it easier for people to teach it new skills. It has a head, eyes, eyelids, lips, arms, hands, waist, and legs. It uses 53 motors to move. It has all kinds of sensors to help it understand its environment and its body position. It has two cameras to see, two microphones to hear, and a smart “skin” that lets it sense if someone is touching it. It has sensors at all joints. The iCub also has accelerometers and gyroscopes that indicate how fast it moves or turns. It also has a couple of inertia sensors. Still with all the above skills, the iCub is still not able to perform most of the tasks that a human can perform yet, over the time of its “existence” it has been learning new skills and hopefully, iCub will help researchers share what they have learned and keep its development moving forward.
A strong and robust robot must have five principal characteristics:
- The capacity to coordinate its actuator and sensors to actively explore the environment.
- Behaviour on micro and macro levels.
- Bi-directional interaction between the system and the environment.
- Bi-direction communication between the system and the other agents.
- Understanding of the physics of the environment.
Embodiment and the interaction between the intelligent system and the environment are crucial to improve the robustness and decrease the errors originated by the fact that real world is unpredictable. The more the embodied intelligent system is integrated and connected to the world in the same way as an animal, the stronger and more capable to respond to problems it is.
3.4. Situatedness
The previous section outlined how much physical interaction between a programmed robot and its environment is crucial to a robust embodied intelligence. But if taking the point of view of human beings enables to build more robust machines, one can wonder how a robot can interact if it cannot localize accurately in the real world? Interaction is indeed achieved in particular through situatedness, or the ability to adapt to a changing and complex environment ([57]). Of course, one could argue that every machine is situated in a way. However, the complexity of evolving with a body in a changing environment is characteristic of embodied robotics, and that is why robustness relies on situatedness.
The development of the best situated robot possible hinges on the level of unpredictability in an environment. If mankind is quite accustomed to the environment of Earth, the outer space and other planets are still very mysterious places, where forces applied to a robot are not easily computable. Thus, what better example than a space robotic assistant and explorer to develop the most efficient situated robot possible?
Robonaut 2 is an anthropomorphic robot used in the ISS (International Space Station) to assist crew members, by relieving them from stressful activities, and eventually helping them with EVA (Extravehicular Activities) ([29]). The objective of Robonaut 2, designed in partnership with the industrial company General Motors, is to complete high-dexterity tasks ([74]). Robonaut 2 is a good example of an Embodied Intelligence application, mostly because of its situatedness ability that enables it to execute complex routines. Indeed, it uses a motion planning system able to respect high-level requirements and constraints ([47]). The following example addresses the simple task of opening a valve aboard the station. This basic movement (which consists in turning the wheel and sliding the door) has to take into account the robot’s body shape, the different degrees of freedom, but also the particular environment of the ISS.
It is considered that each constraint can be represented by a distance between an actuator and a target. This problem can thus be divided into four constraints, represented in Figure 3.4.1:
- (a) The distance of the left foot to the grasp
- (b) The distance of the right foot to the grasp (both feet are attached to the wall or ground of the station)
- (c) The angle of the waist
- (d) The distance between the hand and the grasp (here, the valve)
Figure 3.4.1: Spatial description of the constraints that must be satisfied for Robonaut 2 to open a valve, i.e., F(q) = 0 for all q (from [47])
Q is defined as the configuration space, and F is the constraint function equal to zero when q∈ Q respects the constraint. With these notations, the constrained configuration space X can be defined as a sub-part of the configuration space. It is composed of all the elements q of the configuration space that cancel the constraint function F.
The definition of this constrained configuration space is the starting point of the motion planning algorithm used by Robonaut 2. Its function is to solve a high requirement and constrained motion, and it enables the robot to assess its situational awareness in order to eventually execute a movement adapted to the situation.
This robot also demonstrates situatedness when evolving in an uncontrolled and unique environment, as it creates a real-time 3D model of its surrounding rather than consulting pre-loaded maps. In the future, the key feature of this robot (in addition to assisting astronauts inside and outside the station) will be its ability to explore rough terrains on rocky bodies outside the Earth. This is why situatedness will be its most important skill, as it will have to operate safely in a hazardous – if not completely unknown – terrain.
Robonaut situatedness also enhances its physical interactions with crew members (see Figure 3.4.2).
Figure 3.4.2 : Robonaut 2 uses situatedness to interact with humans nearby (from [5])
Indeed, the robot itself has been designed for interaction, as it uses more than 350 sensors (measuring position, temperature, motor current, force or tension amongst other variables) and 50 motors to evaluate and assess its environment.
Robonaut 2 interaction ability can also be assessed through its capacity of sensing nearby humans and adapt its behaviour accordingly. With the case of an arm for instance, the operator can set a joint torque limit. This limit is useful because the position measurements are computed with a torque control loop. This means that the position determined by the algorithm is computed with the ratio of the commanded torque for a joint to that joint’s torque limit. Consequently, once this limit is reached, the robot will stop resisting to the pressure applied by the person and let it manipulate its arm. This protection ensures the safety of humans around the robot, as it will stop working with very little effort if necessary.
As demonstrated with the example of Robonaut 2, situatedness is an essential feature for embodied intelligence, as it uses the robot’s perception to assess its environment and interact safely through its body. This property is equally applicable to other fields of application, such as research or industry.
As a matter of fact, embodied intelligence is studied nowadays for the purpose of interdisciplinary emotion research ([21]). For such applications in this field of research, situatedness revealed to be crucial, as emotions are often triggered by an external stimulus. The behaviour of an emotional robot, or how it interacts with the environment, will thus depend on its ability to localize itself.
Embodied industrial robots can also use their situatedness ability to enhance their skills. The European project Robominers ([52]), developed to tackle the current rarefaction of mineral resources, is relevant here. Indeed, the purpose of this application is to navigate, operate and perform selective mining by continuously assessing its environment. The robustness of Robominers will completely rely on its situatedness ability, because the better they can localize underground, the more efficiently they will move and find resources. This feature will be achieved through the perception of physical inputs such as touch, conductivity or inertial movements – since a typical GPS system cannot be used underground. Similarly to Robonaut 2, each Robominer will optimize its interactions with the underground environment thanks to its many sensors. The robot will then adapt its behaviour to its surrounding, by constantly updating a map in a relative coordinate frame and adjusting its position in consequence with actuators.
Situatedness is essential for Embodied Intelligence applications, as it enables them to assess their environment. In each field of application, embodied robots use situatedness to interact better with the real world by adapting their behaviours: Robonaut 2 can assist astronauts safely, whereas situated robots are used for research on emotions, and Robominers could revolutionize the exploitation of mining resources. The better an embodied intelligence can adapt to external stimuli, the better its robustness.
3.5. Likeness
Figure 3.5.1: The Uncanny Valley from [60]
Through Embodied Intelligence, scientists are trying to achieve a better interaction between humans and robots. This interaction does not only concern the physical interaction but also the psychological one. This includes the emotions of the machine and, more importantly, the emotions of the user that are triggered by the behaviour of the robot and its physical appearance. Indeed, the goal is to make the user as comfortable as possible when interacting with a machine, raising the question of the likeness of the body. In 1970, Masahiro Mori asked himself “How comfortable do we feel in the presence of something humanlike?” ([60]). Along its research, he proposed the concept of the Uncanny Valley that is illustrated in Figure 3.5.1 and can be interpreted as follows. The more a robot is human like, the more we feel comfortable in its presence. But at a certain level of likeness, the trend reverses itself and we start to feel a repulsion and a feeling of eerie toward the robot. The comfort and discomfort that we feel is amplified by the movement of the robot.
Masahiro Mori initially suggests avoiding the Uncanny Valley by building robots that keep industrial features instead of trying to mimic the human appearance. This could be achieved by showing metal arms, having eyes on a screen such a RAGI, etc.
Understanding this valley is important. Indeed, according to the philosophy of Rodney Brooks, we will live in symbiosis with robots in a near future. But if we want to accept them into the society, we need to feel comfortable around them and to trust them. Thus, knowing the cause of the gap of trust can help scientists and roboticists build the future generation of robots with which we will interact.
The first cause of the Uncanny Valley is the reminder of death and mortality as shown in Figure 3.5.2 with the reference to a corpse or a zombie. This is the original hypothesis raised by Masahiro Mori and studied in details by Karl MacDorman ([54]). This has been tested using a psychological theory, the terror management theory, by showing images of either a human or an android to participants and asking them to describe their feeling. If the reminder of death is unconscious, it raises ethical question regarding the place of the android into the society. However, if this reminder is conscious, it can be considered during the building of the machine or eased by habituating the people to this effect.
Karl MacDorman suggests another explanation to the Uncanny Valley that he calls the “Expectation Violation” ([55]). If an android is humanlike, our expectations (mainly subconscious) toward this robot are similar to the ones we place in other people. But often, the androids are not capable to use the proper reaction in response to a human behaviour. This fail of expectation makes us unable to consider the robot as a living creature or make us feel they are breaking the norm-oriented response. To overcome this problem, the android should be able to recreate the non-verbal language of humans.
Another explanation is proposed by Ramey ([72]). According to his proposal, the Uncanny Valley result from a questioning about categories that are usually differentiated. When exposed to an android that is humanlike enough, we try to put in a category that is neither human nor robot. This forces us to reconsider the definition of humanity and thus reconsidering ourselves. Ramey extend his philosophy to any bio-inspired robot that looks close enough to the host animal.
Figure 3.5.2: The robots of the blog in the Uncanny Valley. Graph modified from [60], Kismet modified from Kismet [93], NAO modified from [97], Kaspar modified from [45], Dolores modified from [30].
The different robots presented in this blog can be ranked on the graph of the Uncanny Valley. For example, Kismet, a robot developed by Rodney Brooke ([17]) is on the leftmost part of the graph. It has some human features such as eyes and ears but cannot be seen as a real being. In the bottom of the Valley, we have the robot Kaspar, introduced in Section 3.1, developed to help autistic children learn social rules and interactions. It has a face with a smile always printed on its visage and hollow eyes. The robot has achieved some interesting results by helping children creating interactions with people and controlling their emotions. Those results could be improved by working on the appearance of the robot. Between those two robots, on the first peak of the graph, we have the robot NAO, also developed to help autistic children, that is presented in Section 4.5. Finally, the robot Dolores from Westworld is in the far right of the graph. It looks exactly like a human and it is trusted by humans.
3.6. Towards a definition of embodied intelligence
Brooks himself never really gave an exact definition of Embodied Intelligence. The field emerged by itself driven by the need to address the issues of its time. But some patterns can be drawn from the above. All characteristics reviewed in the last subsections are essential when designing an embodied intelligent robot. Some are at the base of the field, like the behavioral approach or the importance of the interactions with the environment, and some come more as a “corollary”, like situatedness and likeness – which does not make them dispensable requirements.
Rolf Pfeifer ([69]) attempted the following definition, based on Brooks’ approach:
- Embodied Intelligence: a mechanism (mechanical, biological, electrical…) that learns how to survive in a hostile environment.
This simple definition implies that the intelligence will interact with the environment through its sensors. Those and the actuators of the robot are part of the embodiment, which gave its name to this new field of research and acts as a bridge between the intelligence core and the environment.
One can think of the embodiment as everything taking information for the robot AND everything it can use to modify its behavior. For example, think about yourself as an embodied machine. When you are driving, your eyes and the speed counter, as sensory inputs, are part of your embodiment, but also the wheel, your feet and the pedals, as they serve to modify your direction and speed.
Figure 3.6.1 : Visualization of an Embodied Intelligence and its environment as defined by Pfeifer ([85]).
Janusz Starzyk ([85]) later summed up the researches of his predecessor and managed to bring out the design principles that are inherent to embodied intelligence. Here is a short list summarizing them :
- No creation of a model of the environment. The behaviors will emerge from the direct interactions with it, and create the right links between sensory inputs and motor functions.
- For robustness, there must be a balance between the received information, the processing capability and the complexity of the motor control.
- The design must be cheap and redundant, requiring functionality overlap between some subprocesses.
- The processes must be parallel and loosely-coupled, as it is their interactions and not the processes themselves that will be the factor for intelligence to emerge.
- The robot must either have set goals, or must be able to create goals for itself, often using a subsumption system.
All of these principles have been either applied or stated by Brooks in his work, and for the basis of the field of research. It is easy to see, for some of them, how much they differ from the current, more computational and disembodied, trend in AI. Instead, Embodied intelligence is an engineering and physical approach to design intelligent machines.
Embodied Intelligence 