Robotics

“This book is not about the psychology or cultural anthropology of robotics, interesting as those are. I am an engineer and roboticist, so I confine myself firmly to the technology and application of real physical robots. […] robotics is the study of the design, application, and use of robots, and that is precisely what this Very Short Introduction is about: what robots do and what roboticists do.”

…

The above quote is from the book‘s preface; the book is quite decent and occasionally really quite fascinating. Below I have added some sample quotes and links to topics/stuff covered in the book.

…

“Some of all of […] five functions – sensing, signalling, moving, intelligence, and energy, integrated into a body – are present in all robots. The actual sensors, motors, and behaviours designed into a particular robot body shape depend on the job that robot is designed to do. […] A robot is: 1. an artificial device that can sense its environment and purposefully act on or in that environment; 2. an embodied artificial intelligence; or 3. a machine that can autonomously carry out useful work. […] Many real-world robots […] are not autonomous but remotely operated by humans. […] These are also known as tele-operated robots. […] From a robot design point of view, the huge advantage of tele-operated robots is that the human in the loop provides the robot’s ‘intelligence’. One of the most difficult problems in robotics — the design of the robot’s artificial intelligence — is therefore solved, so it’s not surprising that so many real-world robots are tele-operated. The fact that tele-operated robots alleviate the problem of AI design should not fool us into making the mistake of thinking that tele-operated robots are not sophisticated — they are. […] counter-intuitively, autonomous robots are often simpler than tele-operated robots […] When roboticists talk about autonomous robots they normally mean robots that decide what to do next entirely without human intervention or control. We need to be careful here because they are not talking about true autonomy, in the sense that you or I would regard ourselves as self-determining individuals, but what I would call ‘control autonomy’. By control autonomy I mean that the robot can undertake its task, or mission, without human intervention, but that mission is still programmed or commanded by a human. In fact, there are very few robots in use in the real world that are autonomous even in this limited sense. […] It is helpful to think about a spectrum of robot autonomy, from remotely operated at one end (no autonomy) to fully autonomous at the other. We can then place robots on this spectrum according to their degree of autonomy. […] On a scale of autonomy, a robot that can react on its own in response to its sensors is highly autonomous. A robot that cannot react, perhaps because it doesn’t have any sensors, is not.”

“It is […] important to note that autonomy and intelligence are not the same thing. A robot can be autonomous but not very smart, like a robot vacuum cleaner. […] A robot vacuum cleaner has a small number of preprogrammed (i.e. instinctive) behaviours and is not capable of any kind of learning […] These are characteristics we would associate with very simple animals. […] When roboticists describe a robot as intelligent, what they mean is ‘a robot that behaves, in some limited sense, as if it were intelligent’. The words as if are important here. […] There are basically two ways in which we can make a robot behave as if it is more intelligent: 1. preprogram a larger number of (instinctive) behaviours; and/or 2. design the robot so that it can learn and therefore develop and grow its own intelligence. The first of these approaches is fine, providing that we know everything there is to know about what the robot must do and all of the situations it will have to respond to while it is working. Typically we can only do this if we design both the robot and its operational environment. […] For unstructured environments, the first approach to robot intelligence above is infeasible simply because it’s impossible to anticipate every possible situation a robot might encounter, especially if it has to interact with humans. The only solution is to design a robot so that it can learn, either from its own experience or from humans or other robots, and therefore adapt and develop its own intelligence: in effect, grow its behavioural repertoire to be able to respond appropriately to more and more situations. This brings us to the subject of learning robots […] robot learning or, more generally, ‘machine learning’ — a branch of AI — has proven to be very much harder than was expected in the early days of Artificial Intelligence.”

“Robot arms on an assembly line are typically programmed to go through a fixed sequence of moves over and over again, for instance spot-welding car body panels, or spray-painting the complete car. These robots are therefore not intelligent. In fact, they often have no exteroceptive sensors at all. […] when we see an assembly line with multiple robot arms positioned on either side along a line, we need to understand that the robots are part of an integrated automated manufacturing system, in which each robot and the line itself have to be carefully programmed in order to coordinate and choreograph the whole operation. […] An important characteristic of assembly-line robots is that they require the working environment to be designed for and around them, i.e. a structured environment. They also need that working environment to be absolutely predictable and repeatable. […] Robot arms either need to be painstakingly programmed, so that the precise movement required of each joint is worked out and coded into a set of instructions for the robot arm or, more often (and rather more easily), ‘taught’ by a human using a control pad to move its end-effector (hand) to the required positions in the robot’s workspace. The robot then memorizes the set of joint movements so that they can be replayed (over and over again). The human operator teaching the robot controls the trajectory, i.e. the path the robot arm’s end-effector follows as it moves through its 3D workspace, and a set of mathematical equations called the ‘inverse kinematics’ converts the trajectory into a set of individual joint movements. Using this approach, it is relatively easy to teach a robot arm to pick up an object and move it smoothly to somewhere else in its workspace while keeping the object level […]. However […] most real-world robot arms are unable to sense the weight of the object and automatically adjust accordingly. They are simply designed with stiff enough joints and strong enough motors that, whatever the weight of the object (providing it’s within the robot’s design limits), it can be lifted, moved, and placed with equal precision. […] The robot arm and gripper are a foundational technology in robotics. Not only are they extremely important as […] industrial assembly-line robot[s], but they have become a ‘component’ in many areas of robotics.”

“Planetary rovers are tele-operated mobile robots that present the designer and operator with a number of very difficult challenges. One challenge is power: a planetary rover needs to be energetically self-sufficient for the lifetime of its mission, and must either be launched with a power source or — as in the case of the Mars rovers — fitted with solar panels capable of recharging the rover’s on-board batteries. Another challenge is dependability. Any mechanical fault is likely to mean the end of the rover’s mission, so it needs to be designed and built to exceptional standards of reliability and fail-safety, so that if parts of the rover should fail, the robot can still operate, albeit with reduced functionality. Extremes of temperature are also a problem […] But the greatest challenge is communication. With a round-trip signal delay time of twenty minutes to Mars and back, tele-operating the rover in real time is impossible. If the rover is moving and its human operator in the command centre on Earth reacts to an obstacle, it’s likely to be already too late; the robot will have hit the obstacle by the time the command signal to turn reaches the rover. An obvious answer to this problem would seem to be to give the rover a degree of autonomy so that it could, for instance, plan a path to a rock or feature of interest — while avoiding obstacles — then, when it arrives at the point of interest, call home and wait. Although path-planning algorithms capable of this level of autonomy have been well developed, the risk of a failure of the algorithm (and hence perhaps the whole mission) is deemed so high that in practice the rovers are manually tele-operated, at very low speed, with each manual manoeuvre carefully planned. When one also takes into account the fact that the Mars rovers are contactable only for a three-hour window per Martian day, a traverse of 100 metres will typically take up one day of operation at an average speed of 30 metres per hour.”

“The realization that the behaviour of an autonomous robot is an emergent property of its interactions with the world has important and far-reaching consequences for the way we design autonomous robots. […] when we design robots, and especially when we come to decide what behaviours to programme the robot’s AI with, we cannot think about the robot on its own. We must take into account every detail of the robot’s working environment. […] Like all machines, robots need power. For fixed robots, like the robot arms used for manufacture, power isn’t a problem because the robot is connected to the electrical mains supply. But for mobile robots power is a huge problem because mobile robots need to carry their energy supply around with them, with problems of both the size and weight of the batteries and, more seriously, how to recharge those batteries when they run out. For autonomous robots, the problem is acute because a robot cannot be said to be truly autonomous unless it has energy autonomy as well as computational autonomy; there seems little point in building a smart robot that ‘dies’ when its battery runs out. […] Localization is a[nother] major problem in mobile robotics; in other words, how does a robot know where it is, in 2D or 3D space. […] [One] type of robot learning is called reinforcement learning. […] it is a kind of conditioned learning. If a robot is able to try out several different behaviours, test the success or failure of each behaviour, then ‘reinforce’ the successful behaviours, it is said to have reinforcement learning. Although this sounds straightforward in principle, it is not. It assumes, first, that a robot has at least one successful behaviour in its list of behaviours to try out, and second, that it can test the benefit of each behaviour — in other words, that the behaviour has an immediate measurable reward. If a robot has to try every possible behaviour or if the rewards are delayed, then this kind of so-called ‘unsupervised’ individual robot learning is very slow.”

“A robot is described as humanoid if it has a shape or structure that to some degree mimics the human form. […] A small subset of humanoid robots […] attempt a greater degree of fidelity to the human form and appearance, and these are referred to as android. […] It is a recurring theme of this book that robot intelligence technology lags behind robot mechatronics – and nowhere is the mismatch between the two so starkly evident as it is in android robots. The problem is that if a robot looks convincingly human, then we (not unreasonably) expect it to behave like a human. For this reason whole-body android robots are, at the time of writing, disappointing. […] It is important not to overstate the case for humanoid robots. Without doubt, many potential applications of robots in human work- or living spaces would be better served by non-humanoid robots. The humanoid robot to use human tools argument doesn’t make sense if the job can be done autonomously. It would be absurd, for instance, to design a humanoid robot in order to operate a vacuum cleaner designed for humans. Similarly, if we want a driverless car, it doesn’t make sense to build a humanoid robot that sits in the driver’s seat. It seems that the case for humanoid robots is strongest when the robots are required to work alongside, learn from, and interact closely with humans. […] One of the most compelling reasons why robots should be humanoid is for those applications in which the robot has to interact with humans, work in human workspaces, and use tools or devices designed for humans.”

“…to put it bluntly, sex with a robot might not be safe. As soon as a robot has motors and moving parts, then assuring the safety of human-robot interaction becomes a difficult problem and if that interaction is intimate, the consequences of a mechanical or control systems failure could be serious.”

“All of the potential applications of humanoid robots […] have one thing in common: close interaction between human and robot. The nature of that interaction will be characterized by close proximity and communication via natural human interfaces – speech, gesture, and body language. Human and robot may or may not need to come into physical contact, but even when direct contact is not required they will still need to be within each other’s body space. It follows that robot safety, dependability, and trustworthiness are major issues for the robot designer. […] making a robot safe isn’t the same as making it trustworthy. One person trusts another if, generally speaking, that person is reliable and does what they say they will. So if I were to provide a robot that helps to look after your grandmother and I claim that it is perfectly safe — that it’s been designed to cover every risk or hazard — would you trust it? The answer is probably not. Trust in robots, just as in humans, has to be earned. […for more on these topics, see this post – US] […] trustworthiness cannot just be designed into the robot — it has to be earned by use and by experience. Consider a robot intended to fetch drinks for an elderly person. Imagine that the person calls for a glass of water. The robot then needs to fetch the drink, which may well require the robot to find a glass and fill it with water. Those tasks require sensing, dexterity, and physical manipulation, but they are problems that can be solved with current technology. The problem of trust arises when the robot brings the glass of water to the human. How does the robot give the glass to the human? If the robot has an arm so that it can hold out the glass in the same way a human would, how would the robot know when to let go? The robot clearly needs sensors in order to see and feel when the human has taken hold of the glass. The physical process of a robot handing something to a person is fraught with difficulty. Imagine, for instance, that the robot holds out its arm with the glass but the human can’t reach the glass. How does the robot decide where and how far it would be safe to bring its arm toward the person? What if the human takes hold of the glass but then the glass slips; does the robot let it fall or should it — as a human would — renew its grip on the glass? At what point would the robot decide the transaction has failed: it can’t give the glass of water to the person, or they won’t take it; perhaps they are asleep, or simply forgotten they wanted a glass of water, or confused. How does the robot sense that it should give up and perhaps call for assistance? These are difficult problems in robot cognition. Until they are solved, it’s doubtful we could trust a robot sufficiently well to do even a seemingly simple thing like handing over a glass of water.”

“The fundamental problem with Asimov’s laws of robotics, or any similar construction, is that they require the robot to make judgments. […] they assume that the robot is capable of some level of moral agency. […] No robot that we can currently build, or will build in the foreseeable future, is ‘intelligent’ enough to be able to even recognize, let alone make, these kinds of choices. […] Most roboticists agree that for the foreseeable future robots cannot be ethical, moral agents. […] precisely because, as we have seen, present-day ‘intelligent’ robots are not very intelligent, there is a danger of a gap between what robot users believe those robots to be capable of and what they are actually capable of. Given humans’ propensity to anthropomorphize and form emotional attachments to machines, there is clearly a danger that such vulnerabilities could be either unwittingly or deliberately exploited. Although robots cannot be ethical, roboticists should be.”

“In robotics research, the simulator has become an essential tool of the roboticist’s trade. The reason for this is that designing, building, and testing successive versions of real robots is both expensive and time-consuming, and if part of that work can be undertaken in the virtual rather than the real world, development times can be shortened, and the chances of a robot that works first time substantially improved. A robot simulator has three essential features. First, it must provide a virtual world. Second, it must offer a facility for creating a virtual model of the real robot. And third, it must allow the robot’s controller to be installed and ‘run’ on the virtual robot in the virtual world; the controller then determines how the robot behaves when running in the simulator. The simulator should also provide a visualization of the virtual world and simulated robots in it so that the designer can see what’s going on. […] These are difficult challenges for developers of robot simulators.”

“The next big step in miniaturization […] requires the solution of hugely difficult problems and, in all likelihood, the use of exotic approaches to design and fabrication. […] It is impossible to shrink mechanical and electrical components, or MEMS devices, in order to reduce total robot size to a few micrometres. In any event, the physics of locomotion through a fluid changes at the microscale and simply shrinking mechanical components from macro to micro — even if it were possible — would fail to address this problem. A radical approach is to leave behind conventional materials and components and move to a bioengineered approach in which natural bacteria are modified by adding artificial components. The result is a hybrid of artificial and natural (biological) components. The bacterium has many desirable properties for a microbot. By selecting a bacterium with a flagellum, we have locomotion perfectly suited to the medium. […] Another hugely desirable characteristic is that the bacteria are able to naturally scavenge for energy, thus avoiding the otherwise serious problem of powering the microbots. […] Whatever technology is used to create the microbots, huge problems would have to be overcome before a swarm of medical microbots could become a practical reality. The first is technical: how do surgeons or medical technicians reliably control and monitor the swarm while it’s working inside the body? Or, assuming we can give the microbots sufficient intelligence and autonomy (also a very difficult challenge), do we forgo precise control and human intervention altogether by giving the robots the swarm intelligence to be able to do the job, i.e. find the problem, fix it, then exit? […] these questions bring us to what would undoubtedly represent the greatest challenge: validating the swarm of medical microbots as effective, dependable, and above all safe, then gaining approval and public acceptance for its use. […] Do we treat the validation of the medical microbot swarm as an engineering problem, and attempt to apply the same kinds of methods we would use to validate safety-critical systems such as air traffic control systems? Or do we instead regard the medical microbot swarm as a drug and validate it with conventional and (by and large) trusted processes, including clinical trials, leading to approval and licensing for use? My suspicion is that we will need a new combination of both approaches.”

…

Links:

E-puck mobile robot.
Jacques de Vaucanson’s Digesting Duck.
Cybernetics.
Alan Turing. W. Ross Ashby. Norbert Wiener. Warren McCulloch. William Grey Walter.
Turtle (robot).
Industrial robot. Mechanical arm. Robotic arm. Robot end effector.
Automated guided vehicle.
Remotely operated vehicle. Unmanned aerial vehicle. Remotely operated underwater vehicle. Wheelbarrow (robot).
Robot-assisted surgery.
Lego Mindstorms NXT. NXT Intelligent Brick.
Biomimetic robots.
Artificial life.
Braitenberg vehicle.
Shakey the robot. Sense-Plan-Act. Rodney Brooks. A robust layered control system for a mobile robot.
Toto the robot.
Slugbot. Ecobot. Microbial fuel cell.
Scratchbot.
Simultaneous localization and mapping (SLAM).
Programming by demonstration.
Evolutionary algorithm.
NASA Robonaut. BERT 2. Kismet (robot). Jules (robot). Frubber. Uncanny valley.
AIBO. Paro.
Cronos Robot. ECCEROBOT.
Swarm robotics. S-bot mobile robot. Swarmanoid project.
Artificial neural network.
Symbrion.
Webots.
Kilobot.
Microelectromechanical systems. I-SWARM project.
ALICE (Artificial Linguistic Internet Computer Entity). BINA 48 (Breakthrough Intelligence via Neural Architecture 48).