Flight of the Robotic Bees / Robert Wood, Radhika Nagpal and Gu-Yeon Wei

Thousands of robotic insects will take off with a common goal

Not long ago, a mysterious epidemic called "colony collapse disorder" (CCD) began to wipe out honey bee hives in the United States. These bees are responsible for most of the commercial pollination activity carried out in the US, and their loss has raised concerns that agriculture will also be affected. In 2009, we, the three authors and our colleagues at Harvard University and Northeastern University, began to think seriously about what it would take to create a robotic bee colony. We wondered if mechanical bees could reproduce not only the behavior of a single bee, but also the unique behavior that occurs when thousands of bees interact with each other. We therefore created the first "RoboBees", flying robots that are the size of a bee, and we are working on developing ways to make thousands of them cooperate like in a real beehive.

At first glance, such a task seems almost impossible. The bees have been shaped over millions of years of evolution into amazing flying machines. The tiny creature is able to fly for hours, maintain stability during gusts of wind, search for flowers and evade predators. Try doing this with a coin-sized robot!

And what about the hive? It seems that there is no supervisor and no centralized authority in the bee colony. Nevertheless, tens of thousands of honey bees in the colony divide the work among themselves wisely and manage to perform essential tasks for the existence of the entire hive. When the hive needs more pollen, more bees go out to collect it; When the hive requires maintenance, the bees stay at home. And when something goes wrong, for example when the queen dies unexpectedly, the bees quickly adapt to the changed circumstances. If there is no leader, how is such a large colony able to make these complex decisions quickly and without chaos caused by communication problems?

A robotic beehive could help agriculture by incubating flowers, but that's just one of many possible uses. In fact, small, agile, simple and inexpensive robots will be able to perform many operations with much greater efficiency than a few robots with high capabilities can perform. For example, think of a rescue worker equipped with a box of a thousand bee robots that weighs less than one kilogram. These beaver robots can be released at a disaster site to search for the heat, sound, or carbon dioxide signature of survivors. It is enough for only three robots out of a thousand to succeed in their mission, and this will be considered a success for the entire swarm. This is something that cannot be said about the current generation of rescue robots, which cost hundreds of thousands of dollars per unit.

However, a robotic bee colony poses a huge number of technological challenges. The size of the robots should be a few centimeters from end to end, and weigh about half a gram - about the weight of the lightest autonomous aircraft in the world. This tiny package should include the flight system, the vision system as well as the electronic brain, and the control means that will determine the way the bee interacts with its hive mates. All of these goals are coming closer to reach thanks to advances in materials science, sensor technology, and computing architecture.

body and flight

The most obvious challenge in creating a tiny flying robot is how to make it fly. Unfortunately, the progress achieved in the field of miniaturizing robots in the last ten years does not help us too much, because the small dimensions of the devorobot change the nature of the forces that act on it. Surface forces, such as friction, become more influential than volume-related forces, such as gravity and inertia. The problem of scale renders most of the typical mechanical engineer's toolbox ineffective, including bearings, gears, and electromagnetic motors. These components are very common in larger robots, but are not efficient enough for Devorobot.

Instead of motors and spinning gears, we designed the Dobrobot with an anatomy very similar to that of a flying insect: flapping wings powered (in this case) by artificial muscles [see frame above]. The muscle system we created uses separate "muscles" for propulsion and control. Relatively large and powerful motion generators (actuators) cause oscillations in the wing and chest system to create the flutter, while smaller motion generators adjust the movements of the wings to create a turning moment for control and maneuvering purposes. The two motion generators act on the joint that connects the wing to the body.

The artificial muscle is made of piezoelectric materials, which contract when an electrical voltage is applied to them. Such motion generators have some disadvantages, for example, they require high voltage and are fragile, but this is one of the cases where the physics of the small scale works in our favor. The smaller these motion generators are, the faster they try to operate, and since the amount of work needed per cycle (per unit mass) remains more or less constant, faster flailing creates higher power. In fact, these muscles produce power comparable to that of similarly sized insect muscles.

In recent years, we have conducted experiments with dozens of different configurations of motion generators and joints. And since it would be necessary to mass-produce thousands of such bees, we examined the ease of construction in each configuration.

The most successful artificial muscle models we've been able to build so far are made from a three-layer sandwich: rigid sheets on top and bottom, and a thin polymer sheet in the middle. We create the flexible joints by removing material from the outer layers through etching, so that the polymer in the middle is free to bend [see frame on the previous page].

We've made great progress in building a bee-sized robot, but we're still trying to figure out the best way to power it. To meet the high energy requirements of small-scale flight, most of the mass of the robotic bee needs to be concentrated in the main motion generator and the power unit (for the purpose of the discussion we will refer to it as a battery, although we are also looking at using a tiny solid-oxide type fuel cell.) The question of energy has been revealed to a degree Somewhat as a catch 22: the larger the power unit, the more energy it stores, but requires a larger propulsion system to handle the excess weight, which in turn requires an even larger energy source.

We have not yet been able to fly a devorobot with an independent energy source, yet we have demonstrated how a robot weighing 100 milligrams produces enough thrust to take off (when it is connected by wires to an external energy source). The devorobot was also able to stabilize itself using a combination of active and passive mechanisms. Given the energy density of the best modern batteries and the efficiency of all the robot's components, we estimate that the maximum flight time will not exceed only a few tens of seconds. In order to extend this flight time, we are working to reduce the mass and maximize the efficiency of each component of the robot's body.

brain and navigation

The energy is not the only reason why the Deborobot is still connected by a cable. Another problem that we have not yet solved is that of the carried "brain": an independent bee robot in the field will have to scan the environment, decide what is the preferred course of action and control the flight mechanisms. In the lab we use an improvised solution of external electronics, but in the field, the devorobot will need its own "brain".

At the level of the higher processes, this brain constitutes an intelligent unit that is not only responsible for controlling the single beaver robot, but also for managing its interaction with other beaver robots in the colony. We build the brain in layers: sensors for deciphering the physical environment, an electronic nervous system that handles basic control operations, and a programmable electronic "cortex" that makes decisions at a high level. In the first step, we tried to design a subsystem of this brain, which would enable independent flight. This challenge requires a tight control circuit that includes sensors, signal processors and movement of body parts.

To determine which sensors to use and how to wire the robot's brain, we once again turned to nature. Flies (and other animals) use two main types of sensors to navigate the world: proprioceptive sensors that give the flyer information about himself - how fast his wings are moving, for example, or how much electricity is left in the battery, and extroceptive sensors that provide information about the outside world.

Modern technology offers satellite navigation, acceleration sensors, and multi-axis gyroscopes, but such sensors are often heavy, power-hungry, or both, and therefore unsuitable for such use. That is why we are testing an electronic vision system similar to that of real bees: a system that analyzes "optical flow", which is the visible movement of objects in the visual field of an image sensor. Imagine the view from the passenger window of a car: nearby objects seem to pass quickly through the field of vision, while distant objects move slowly. A vision system that utilizes this information is able to create a three-dimensional representation of the environment, even if it is equipped with only a small and simple image sensor.

However, the brain of the devorobot will have to be strong enough to process the stream of data coming from the image sensors, and make the appropriate control decisions to drive the movement generator in its body. Here too, even the advanced components available in the market will not help us. Therefore, we are looking at a new type of computing architecture for the devorobot's brain, which combines multipurpose computing with dedicated circuits known as "hardware accelerators". Unlike multipurpose processors that are capable of performing many tasks, such as running ordinary home computers, hardware accelerators are circuit blocks that have been specifically and precisely adapted to perform one and only one task - but perform it correctly. We will use hardware accelerators to perform fast real-time calculations, necessary for the stable flight feedback control loop, within the limited energy budget.

A major challenge was to determine what compromises we could accept. For example, we would like to use a high-resolution camera, but a large number of pixels requires large image sensors and additional computing power to process the images. So what is the right balance?

To help solve such questions, we have developed a special experimental chamber. We connect the body of a devorobot to a multi-axis force and torque sensor and let it flap its wings in an attempt to fly. On the cell walls we project images of the physical environment in which the beaver robot will fly. In this way it is possible to examine how the prototypes of the vision system, the brain and the body work together and navigate the world.

Of course, flight control is just the beginning. At the same time, we are researching other types of sensors, which will allow Devorobot to perform unique tasks - for example, finding a person trapped among the rubble of an earthquake.

Unfortunately, one ability we don't expect these bees of ours have is direct bee-to-bee communication. The energy costs associated with wireless communication are too great. Nevertheless, this does not mean that the bees will act alone.

Colony and communication

A single devorobot will be tiny and limited in relation to the world in which we hope it will operate, and the energy and weight limitations require us to be content with each devorobot being able to carry limited hardware for sensing and communication. Therefore, apart from the research that deals with the body and brain of the devorobot, we must also understand how to build a colony. Like the honey bees, one robot bee will not be able to achieve much, but with the help of group behavior, a hive of bee robots will be able to explore large areas, combine multiple simple observations into a meaningful picture, divide the work efficiently and succeed even when some of the bees fail. Swarms of small and nimble robots, none of which is essential on its own, will be able to perform a variety of tasks, for example pollinating flowers, or search and rescue tasks in disaster events - that individual robots are unable to do.

Since the early 90s, computer scientists studying a field called "swarm intelligence" have created many powerful algorithms for coordinating actions inspired by social insects, ranging from coordinated search strategies to intelligent division of labor. However, despite these algorithms, swarms of robots cannot be managed in the same way as a single robot is managed.

First, when there are thousands of details, programming and inference at the level of detail is impractical, just as it is impossible to require the average software developer to write the instructions for every physical bit in the computer. Instead, similar to how compilers take instructions in a human-readable programming language and translate them into the zeros and ones that govern the operation of the single transistors inside a microprocessor, we need a way to program the colony as a higher-level, abstract assembly, so that the general instructions are translated into behavioral programs of Details. We need a programming language for colonies.

What language is suitable to describe the activity of a beehive, and what we want to accomplish with the help of our colony of bee robots? There is currently no simple answer, but as a start we have developed two abstract languages. In the "Karma" language, you can create a flow chart of tasks that the colony needs to perform. This chart includes links, which represent conditions for triggering new tasks. The Karma system uses information returned from the individuals to adjust the allocation of resources to tasks, in a way that mimics the role of the beehive in real bee colonies.

Another approach, known as optRAD, sees the colony of flying robots as a liquid that pulsates in the surrounding space. Each individual Devorobot uses a probabilistic algorithm to determine whether it will perform a task, depending on the current state of the environment. Seeing the system as a fluid allows optRAD to think at a high level about the expected results and adapt its behavior to new circumstances.

We need to learn a lot more about building and operating a large robot colony, which includes not tens or hundreds but thousands of independent robots, a number far greater than the number of human operators. With such numbers, even running each robot at the individual level is impractical. Imagine that each robot has an on and off switch: if activating each robot takes five seconds, activating a thousand will require almost an hour and a half. Similar limitations apply to everything from cost to maintenance. Every robot must be cheap, easy to build and simple to operate at the collective level. The ambition is that each operation will be scalable, which will require a fixed time, which does not increase (or at least increases very slowly) as the collective grows.

These challenges led us to create the "Kilobot" system - a collective of hundreds of robots, each as wide as a half shekel coin, that move through vibrations and communicate with similar kilobots in their environment. We can use this group to test the effectiveness of our programming languages ​​and mathematical models for emergent behavior. After all, without experimenting with real hardware, there is no way we will understand the emergent behavior of physical systems.

The collective can be used to test many group behaviors that we would like to eventually achieve with the beaver robot colony. For example, we could ask the group to look for targets in the environment, and as soon as a certain kilobot finds a target, it will report its location to the group. We also made the Kilobot design open source for teams interested in creating their own robots. It is also possible to purchase ready-made kilobots from the educational robot company K-team. We hope that such standardized robotics kits will help bring forth new ideas and encourage collaborative advances in science, such that individual groups are unable to achieve - because ultimately, we too rely on the power of the collective to be more than the sum of our parts.

Future

Although we have made great progress, there is still much work to be done. In our opinion, within a few years we will see flying bee robots in well-controlled laboratory conditions. Five to ten years later, you may see them in mass use.

In 1989 the renowned roboticist Rodney Brooks wrote an article about the advantages of small robots in space exploration. The title of the article was: "Fast, cheap and uncontrolled - a robotic invasion of the solar system". This is of course a pun on the well-known saying of engineers regarding consumer products, that they can be characterized by two, and only two, of these titles: fast, cheap and reliable. When there are many details, the failure of one of them is negligible.

Brooks's interpretation of this robotics concept turned out to be prophetic. If you can create many simple details that work effectively together, who cares if a few fail here and there? The only way to ensure the success of the research robots is to let them occasionally fall from the sky.

__________________________________________________________________________________________________________________________________________________________________

About the authors

Robert Wood (Wood) is a professor of engineering and applied sciences at Harvard University and at that university's Wyss Institute for Biologically Inspired Engineering. In 2012 he received the Alan T. Waterman Award from the American National Science Foundation.

Radhika Nagpal holds the Fred Cowell Chair in Computer Science at Harvard and the Wiss Institute. Her work on collective behavior encompasses the fields of artificial intelligence, robotics and biology.

Go-Yeon Wei holds the Gordon McKay Chair in Electronic Engineering and Computer Science at Harvard. His research interests include a wide range of topics related to energy efficient computing systems.

in brief

Beebots are bee-sized flying robots. Their size requires overcoming a huge variety of physical and computational challenges. In such small dimensions, the motors and bearings available in the market are not efficient enough, so the bees must use specially designed artificial muscles to fly and control the flight.

Also, the tiny bees must think for themselves, with the help of tiny sensors that will process information from the environment and processors that will decide which actions to perform.

artificial intelligence

The colony in action

A colony of thousands of bee robots will have to assign tasks to individuals efficiently, even if its information about the environment is only partial. In this scenario, the hive was tasked with locating and pollinating flower fields. First, several bee robots will scan different areas and return to the hive with information about places where there are blooms. The new information will affect where the workers will fly to next. More robots will be assigned to places where there is more work to be done. The hive-based strategy allows bees to exhibit collective intelligence even when energy constraints limit bee-to-bee communication.

Similar to real bees, the bee robots will work at their best as part of a swarm of thousands of individuals, who will coordinate their actions without relying on a single leader. The hive must be stable at such a level that the group can achieve its goals even if many bees are killed.

More on the subject

Kilobot: A Low Cost Scalable Robot System for Collective Behaviors. Michael Rubenstein, Christian Ahler and Radhika Nagpal in 2012 IEEE International Conference on Robotics and Automation (ICRA), pages 3293-3298; May 14-18, 2012.

Progress on "Pico" Air Vehicles. RJ Wood et al. in International Journal of Robotics Research, Vol. 31, no. 11, pages 1292-1302; September 2012.

YouTube channel of the microrobotics lab at Harvard: www.youtube.com/MicroroboticsLab

Videos and more information:

http://io9.com/dont-get-paranoid-but-there-are-now-insect-sized-flyi-487336929

Facebook