PowerPoint

Some observation on current state of roboticsAlbert YefimovRussian FederationSkolkovo

Good day, Ladies and Gentlemen!My name is Albert Yefimov and I am here to talk on emerging technologies in Robotics. But first, I would like to tell few words on myself, so you will understand where my knowledge comes from. I work for Skolkovo Foundation, which is Russias Federal Development Institution. We dont like when we are called Russias Silicon Valley because we are much smaller only 1500 startups plus few big corporations. But essentially we are multi-disciplinary innovation cluster.

1

In Skolkovo I lead Robotics Center where we pull together 60 projects in robotics. Our major focus is service robotics but I prefer to call it human-oriented robotics. We are not research institution so all our projects are startups with major goal of commercialization of some cool technology. No chance to tell you of even few of them but seeing is better than talking so heres video of one of our really innovative projects.VideoThis is ExoAtlet, the only rehabilitation exoskeleton in Europe going through medical clinical trials. There are competitors in US, Israel, Japan but in Europe it is top level project. This is first commercial version, which is designed to help to paraplegics. It is already on the market, and price is 17 thousands euros.

2

Humans are better with weapons

it is comparatively easy to make computers exhibit adult level performance on intelligence tests or playing checkers, and difficult or impossible to give them the skills of a one-year-old when it comes to perception and mobilityHans Moravec, 1988

Humans are better equipped by nature to handle weapons than robots created by humans. To illustrate current state of robotics I picked up these two videos. VideoOne on the left is showing boys doing standard assembly-disassembly of Kalashnikov assault rifle and on the right you can see a famous Janken robot playing rock-scissors-paper game with 100% winning over human. I think it is best so far performance of robots over humans in terms of movement. It is known as Moravec paradox - it is comparatively easy to make computers exhibit adult level performance on intelligence tests or playing checkers, chess or Go and difficult or impossible to give them the skills of a one-year-old when it comes to perception and mobility.

3

Robot definition is very difficult

Can you give definition of Robot?

Robot is what an Expert call a Robot!

Last day we heard a lot of debate on autonomy. But things are much worse - we can not even agree on what robot is. There are some working definitions of robots for statistical purposes only but they are just to understand what are robots of TODAY. But every time we see an advancement like successful spreading robotics vacuum cleaners we have to adjust our definition nobody calls Rumba a robot-cleaner. It is just a cleaner now. I think reason behind such problems lies in more philosophical area than technology. What is artificial intelligence can be actually defined only after we know for sure what is natural intelligence. So far it is not settled question. So if you would ask Dilbert on what is robot his answer will be something like: robot is whatever an expert in robotics calls a robot.

4

State of Machine LearningIf intelligence was a cake, unsupervised learning would be the cake, supervised learning would be the icing on the cake, and reinforcement learning would be the cherry on the cake. We know how to make the icing and the cherry, but we don't know how to make the cake.

Also there is much to explore in machine learning and artificial intelligence. Here I refer to opinion of a leading scientist, who actually gave birth to deep learning and adjacent technologies Yann Le Cunn. He now leads deep learning laboratory in Facebook. Refering to recent AlphaGo victory over Korean Grand Master in Go Le See dol, Yann Le Cunn wrote:If intelligence was a cake, unsupervised learning would be the cake, supervised learning would be the icing on the cake, and reinforcement learning would be the cherry on the cake. We know how to make the icing and the cherry, but we don't know how to make the cake.This is very fair statement of current research on AI. Even if we take as example this Go game, going deeper we see that AlphaGo is actually based on database from many many human games. These unusual, un-human moves comes from the fact that AlphaGo is also played millions of games against itself, evolutioning on the way. But humans have no time to play so many games to learn but still Korean Grandmaster won a score. This means AlphaGo learns at speed far slower than humans do.

5

Hardware for Artificial BrainsHypothetical Full AI would need ~ 10 exaFLOPS, ~500 MW

Cray Titan ~18 pFLOPS

Tianhe-2 ~34 pFLOPS, ~25 MW

Cray-2 (1985)

iPhone 5 (2012)2.7 times more powerful27 years

What is needed for truly Artificial Intelligence? Hardware. There is a race between industrially developed nations to build fastest supercomputer. Cray Titan is now 18 petaFLOPS, where petaFlOP is 10 in power 15, one quadrillion of operations per second. It is 18 million times faster than iphone 6. However, fastest supercomputer now on Earth is Chinese Tianhe-2, with 34 petaFLOPS performance. In order to crunch numbers this monster need 25 Megawatt of electricity. This energy is enough to give electricity to small town of 25 thousands households. If this energy diverts to Geneva lake, then I guess it would be warm enough to swim just now. According to some researchers, true AI will require at least 10exaFLOP performance which is ten thousand times faster than current supercomputers. Supercomputers are very inefficient beasts and this progress will be reached only with five hundred megawatt of energy. It will be probably nuclear powered. I will remind you that iPhone 5 which many people sill using is 2.7 faster than supercomputer of 1985. But it took 27 years to reach that! Yes, some day we will have iPhones fast and powerful enough for artificial intelligence but it is not a thing which might happen in ten or even twenty years.

6

Cambrian Moment of Modern Robotics

There are many notions on so-called Cambrian moment of modern robotics. I believe there are a lot of diplomats, engineers or colonels in this room but I dont expect there are evolutionary biologists here. Well, approximately half billion years ago number of animals on Earth increased tenfold. Some scientists called it Cambrian explosion. There is one possible explanation for it these creatures got eyes and it helps. They see more pray and started better eating habits. They started to see whom they were actually mating and they gave better offsprings. Some roboticists also claim that nowadays we are going through Cambrian moment of robotics because our robots received sensors eyes. They are very good quality and cheap enough to give them almost every robot.

7

Enabling trends in RoboticsMoors LawDigitalization of everything, including productionBetter BatteryHuge communication bandwidth increaseCloud robotics3 devices/human in 2019Storage is cheap commodity

But besides eyes, there are some more things to mention in regards on emerging roboticsFirst, Moors Law lead to increasing on-board and cloud computing performance. It is not only sensors are important but also what to do with data which we receive from sensors. This increase in computing actually let us fusion of all data to increase robots ability for navigation. One of robots advantage in viewing over humans they might use multispectral vision. But in order to do it we need very efficient hardware and software.Then comes digital maps. Digitalization of everything actually helps robots to move around. Without digital maps GOOGLE car is nothing, that is why it can run well in California but immediately stuck in Siberia.We need better battery. That Boston Dynamics Dog was very noisy beast because used internal combustion engine. Unfortunately, nothing beats gasoline so far in efficiency. But improvements are coming. Improvements in communications made possible cloud computing. We talk about autonomous robots but do we realize that their autonomous capacity are not on their board but somewhere? It might be even thousands of kilometers away from robots. Robots are still teleoperated but not by human operator but artificial intelligence. It is interesting how DARPA explained in their rules of Robotics Challenge why teams can actually helped robots to raise with some penalty in minutes. They say that in actual situation if one robot gone then another one will come with already uploaded history of previous robot and keep going without problem.We have huge increase in number of internet connectivity. In a matter of 4 years every human will have 3 devices connected to internet.Finally, storage become commodity and very cheap. So robots can remember all bad things which done to them by humans. Here I refer to these nasty tricks of our collegues from Boston Dynamics pushing and kicking these poor robots.

8

Human Soldier vs Robot SoldierHumanRobotListening16 -20 000 Hz1 - unfinityViewing160 mlns of colors150 MPTouching13 nmDepends on materialsIf feels cold, hot or hungerunhappyDoesnt careEmotionsA lot ofNot at all

Human soldier and robotics soldier have some advantages and disadvantages over one another. Just some of them are summarized in this slide. Yes, robots can listen much better than humans but our vision is still much more faster because our brain has great capacity to process images in parallel. We have very sensitive skin and feel ridges as small as 13 nanometers. But robots dont care if they are cold, hot or hungry. Of course, we feel sorry that robots dont have emotions. Even upon winning games like Go, robots dont even feel joy of victory over humans.

9

Autonomy in military robotics

Let me say few words on autonomy issue. I dont want to go into definition debates as plenty already were said. But what I would like to point out why autonomy is needed by military. I think need of autonomy is not discussed yet but it is important to see what drives it into technology stack of robotics. And I mean true autonomy ability of robot to act in stochastic environment, but not just act fast. Last one is much discussed here automation.

10

Drivers to full autonomy in military roboticsIt is so easy to jam communications of tele/supervised robots or drones. Examples:

DARPA Robotics Challenge

IranU.S. RQ-170 incident

First driver is that teleoperated robots can be jammed and eventually controlled by adversary. Thus, military believes that some autonomy is needed as crucial capability for such functionality like find way home. Heres couple examples. First, DARPA robotics challenge, where robots were teleoperated but communications were jammed. So robots, participated in the competition had only task level autonomy. It was really difficult for them to solve tasks which are just nothing special for us, humans, like open the door. So if anyone afraid of killer robots just dont open the door! Another example is incident few years ago when Iranians downed American drone by spoofing GPS signal. So autonomy is thought as a way to make robots more reliable and proof from human mistakes or adversary actions.

11

Drivers to full autonomy in military roboticsDramatically increase firepower with fewer personnel

Another reason behind autonomy for military robotics is a drive to project more firepower with fewer people. What if some army is totally outnumbered by adversary? Modern military is moving away from paradigm where 6 people control one drone (like now, predator team is 4-6 people) and going towards situation where one serviceman control many drones in many possible situations. Human is still in control but this control is only for critical decision but not teleoperations.

12

Autonomous systemsSeaGroundAir

Few words on situation in autonomous vehicles in Sea, Ground and Air. Not much time to go into very details but just what I believe important trends.

13

Autonomous robots at Sea

ACTUV Sea HunterAutonomy: Mission levelDuration: 20-120 hSpeed: 1-2 M/sConnection: radio, hydroacousticDeepth: 3-6kmRange: 200-300 kmExamples: Klavesin (Ru), Remus (US), Autosub (GB),Hugin (Nw)

RemoraKlavesinSanpaLevel of Autonomy is very highRange is limitedNo weaponsOnly reconnaissance missionsReliable communications are needed!

Autonomy is a very natural choice for marine robotics (underwater and surface). Specially underwater where it is very difficult to maintain tele operation. Recent level of fully autonomous underwater vehicles of serious grade is 200-300 km with few dozen hours of operations at the depths of 6 km. This might me even more if we could solve communication problem. Hydro modems are very limiting at speed. Submersible should come back to surface in order to upload data to command center.

14

Autonomous ground vehicles

Full Functionality Everywhere

FunctionalityLimitedFullSomewhereEverywhere

Space

Autonomous ground vehicles are following one very interesting trend. We see cars with universal functionality operating in very limited environment. This means, that such cars can driver in fully autonomous mode with no human intervention but on very limited territory. Typical representative of this group is google car. As I stated before, google car is nothing without digital maps. There is opposite approach which all major auto manufacturers stick to It, is called limited functionality but everywhere. They expand functionality incrementally but cars are capable of driving in semi-autonomous mode in every possible location. There is no doubt that at some point both approaches will come to one point. But riding along right one is a sure way to maximize benefits of autonomous cars.

15

Unmanned Aerial VehiclesWeight and power source is the biggest limiting factor

Hydrogen Cells are coming to market very soon

Last point on Aerial Vehicle. Single limiting factor on small UAVs is battery efficiency. Plane type UAVs are good but require catapults to launch, have significant vibrations and can not stay in one place. Copters have no such deficiency except one 15-30 minutes flying time. Just not enough for any serious operation and range limiting 2-4 km. This is about to be changed as within 1-2 years we will see commercial UAV projects based on hydrogen cells. Like this one you can see on the right Russian made UAV might stay in air up to 3 hours with 500g useful load. Startup which is doing this project is going to double its performance in a year.

16

Robotics in Russia

Positive image of robot is a very important issue. We, in Russia, are very lucky that in our cultural heritage there are no evil robots. We love robots in our movies, stories and attitudes. Those of you who would like to see our robotics projects might visit our booth at Innorobo, 24 May in Paris. Or come to Skolkovo we are glad to have guests and always arrange visa for our visitors.

17

Therefore, the more technical the world becomes, the more necessary is the promotion of individual freedom and the individual's awareness of himself as a counterbalance.

Albert Speer

It is much discussed here that critical point if machine is delegated right to kill. But in my opinion critical point is delegating machine right to enforce law over human without any human intervention. We all receive speeding notifications from cameras, but whole process is going without any humans involved in decision of enforcing law upon me.I personally believe that our discussion is very important for all of our states and we shall never forget history lessons. Let me read this quoteTherefore, the more technical the world becomes, the more necessary is the promotion of individual freedom and the individual's awareness of himself as a counterbalance.It sounds very modern, isnt it? But in fact, it belongs to a person who not only did a grave crime but person who is convicted nazi criminal Albert Speer. History lessons which are learned so hard shall be never forgotten and we ought to handle technology issues with great care.Thank you very much!

18