[00:00:05]
>> All right so ocean observation is a very important topic for science and for our life because we care about whether we care about hurricanes so we need to gather data on the ocean and there have been a lot of investments from the National Science Foundation and from the Office of Naval Research in this research topic to how to collect data in the ocean there are actually 2 schools of thought currently and of course they always competing for resources so one school of source sorts come from this idea of let's put a high.

[00:00:41]
High energy power line into the ocean and then we put nodes on these power lines and we who have all these sensors capabilities and then we can collect data in an uninterrupted way for a long time so this is called a static sensor network that collects data in the ocean and it started from this Neptune project run jointly by Canada and the U.S. has been there for a long time and also recently this idea has been picked up worn by China has been constructing static sensor networks just like the U.S. and European people are talking about building one.

[00:01:22]
So this is one idea this is a very nice idea that if you can do this and then you can have a lot of data about the ocean. And another school of thought. Is to use mobile sensor networks to collect data in the ocean and this is something happened last year and received some media attention during the fall asked year there have been about 3 hurricanes forming at the same time in the North Atlantic the U.S. Navy No and I says and also a lot of academic researchers just put gliders into the ocean and try to collect data about these 3 hurricanes at the same time this has never.

[00:02:05]
I've been done before at this scale OK so I'm fortunate to be a very small player in this large team so I helped one of the gliders deployed off the coast of Georgia on this project and of course we collect a lot of data and the stated review in size of a hurricane from the science side that now it's not previously known OK so it's a very nice project so OK so if you compare these 2 schools of thoughts one school of thought of using static sensor networks and the other school of thought of using mobile sensor networks from a robotics point of deal.

[00:02:41]
My opinion is they've actually provide 2 different types of data streams so the static sensor networks provide what I call our own data stream so what is an order in data stream it means the data stream the location of the data stream doesn't change and you've got time series data at every location you put a sensor on.

[00:03:00]
And then with mobile phones or networks you actually get what I call a look around the latest train where the data series not only depends on time but also depends on space on the location so space and time are coupled in these data streams of course this long running data streams are more relevant to roboticists like us OK so if you compare this to School of a sought.

[00:03:24]
The other in data streams are easier for math constructs and you can pretty much use them to construct Mab without much work comes naturally however the drawback of this idea is you need lots of censors. In order to actually. Cover an area and build that map. So on the other hand running data streams only need a small amount of mobile sensors so long as you can move around and then you can collect this data however the data they collect needs to be carefully processed before you can use it in other words we need to simulate that into a math OK So this is related to the slam ideas that we work with on the robotics side.

[00:04:11]
All right so over the last 20 years because of the advancements in marine robotics technology we now have a school of different kinds of Marine robots that can go out there in the ocean and collect the data for us OK so here I put 3 different ones that are popular so we have this underwater glider that can actually stay in the ocean for weeks or months or even years now and then we have this wave glider that harvest the wave energy and can stay on a certain surface for for a very long time on the other hand we also have this fast moving Autonomy's underwater vehicles and they're actually driven by pollers and they consume energy in the move pretty fast OK but they cannot last very long.

[00:04:59]
So from a data stream point of deal these underwater robots OK Well Marine robots can be viewed as data collectors and their lives around in a to collectors live because they're moving but they're also different from the ground and data collectors that have previously been used by oceanographers because these are the London data structure previously used by oceanographers the drift the ocean flow OK they don't have control motion Well the Marine robots have controlled motion so actually I highlighted this word Contro I think this is this makes a huge difference for data collection for ocean size.

[00:05:40]
So the idea is to use robots as a terminal to provide control of Greneda streams. This has made a big difference in ocean science. So in order to make this data collectors control it then we need to have autonomy structures wrapped around it so that we can autonomously guided to that these are place to collect data so that is actually exactly what we did things about 15 years ago so every one of this robot has a local feedback control loop and then wrap around it a another feedback control to achieve autonomy so that in this feedback controllable we do is we actually collect all the observation data including the data from the robot OK And then we pass data to a social model and this social model is simulated data so you can think of things of a simulation as math construction OK So uses data to construct a map and then make predictions for future states into the ocean and then this state predicted states are passed back to the robot and the robot will use this predicts state to do pass planning.

[00:06:51]
And then this you repeat this process and becomes a feedback loop wrapped around the robot and make the robot go autonomous lead to collect the data so this is not this is nothing new OK Nothing surprising but there are 2 blocks in this movie that are very important the 1st block is a data simulation BLOCK Why is it important because it's actually converting converting the low ground in data stream into an all or a map.

[00:07:19]
And that conversion takes time. And then another block that is important is this block called planning control imagine it doesn't operate this thing it takes all or in data map and convert it back to the ground in beta. OK so you've got this 2 blocks of things that doing the conversion so why these 2 blocks takes time to compute So this is one slide explanation of data simulation on the lab you have a small number of data collected by a bunch of robots on the right hand side is the oceans that you want to estimate you can compare the size you have very limited data but you want to you want to do a lot with it OK and this takes competition resource and this takes time to compute OK So oftentimes people run a very large that mention common future or comments from a search for this takes a lot of time to compute and it's hard to get it right.

[00:08:16]
So this is one slightly strange in the past landing and this is a state of our past planning against time varying ocean flow so basically you're trying to solve the hermit and Jacoby Bell many question all right and then you just you know propagate this level set forward like solving a P.T. and after you reach this from the starting point of the goal and you go to backtracking you find optimal paths OK we have known this for many years but we cannot compute it OK So we have to find ways to compute this there are still ongoing research in this domain OK that's why these 2 blocks are hard.

[00:08:53]
So over the last 20 years things when you bought it are getting popular and we start to see more and more money robots coming on the market OK and if you look at them their capability of data collections are different so on the left hand side we have this slow moving persistent data collectors that can stay there for a long time but to save energy the euro doesn't move very fast OK So therefore those running data stream they provide are slow and steady and they're basically used for coverage purposes on the right hand side you've got this fast moving a you vs a to them by pollers they move are fast you can go anywhere you want but they don't last very long OK And so therefore their yearly are good for tracking purposes so then the thoughts are how about let's combine the good things on both sides OK So there are actually 2.

[00:09:44]
Approaches in combining the advantages on both sides so the the 1st group Assad is to say well let's build hybrid vehicles so the top row I have a test is made by the U.S. and the civil made by China and they both work similarly so at certain times they can be worked like Sliders and another time they can switch on the propeller and work like a U.V. OK And so so this is this is STATE OF ART OK and then another school of thought is OK let's put a lot of them together a lot of this slow moving persistent data collected together then we can always come in an area and we can also track events OK So on the U.S. side we have a recent demonstration by U.S. Navy to run about $100.00 fighter the same time on the China side the most recent thing I know is about 12 gliders running at the same time OK So these are the state of our approach is to try to compensate for the capabilities of a single Vaghul in terms of data collection OK.

[00:10:50]
Now once you get into this domain. The data collecting behaviors become more complicated then you need to improve our your autonomy structure this is why only Our want to invest into the distributive autonomy and try to engage more development in this type. So started from about 5 years ago we realized that we can actually do a trick to the to the loo of honey structure by putting something called a generic environmental model in between to break this outer feedback loop into 2 smaller feedback loops and again this idea sounds very trivial but it took us a long time to figure out OK You know we're not that smart I have to admit and everything has to be confined by the reality right and the technology vellum level so so I think this is a interesting step forward so if you think about this generic environment a model that actually service data buffers OK so so every time you have a slow data stream and fast data stream in order separated the most naive thing to do is to add a data buffer in between and this is exactly what we did so we added a data buffer in between so then the low over feedback loop will take care of coordinating the motion of the vehicles as a vehicle level of Tommy they can they can run they can handle fast data streams and then the upper level.

[00:12:18]
We can accumulate the data onto certain point and then pass it to the ocean model and then that ocean model processes so this is a much slower data scale OK It turns out this idea works pretty well and if you look at it from an abstract point of view and now there's people talk about Internet of things called computing and computing Actually this is something if you put this into the framework of call computing I call this case of cloud it actually this generic environment motto is actually some it's actually a trick played at the edge computing side OK So at the lower level we can provide process sequential data in a spall small spatial domain with higher resolution in a must faster pace at the cloud level we can process batch data the same time which is a better use of computing resource.

[00:13:12]
And then we worry about a larger spatial domain and longer time horizon of the migration probably. So the this has been what we have been doing in the last 4 or 5 years. And then the recent progress along the sod is OK now we can actually have a distributed of how many architecture.

[00:13:35]
To ask that the genetically genetically environment tomatoes to be installed for every platform OK And then we can ask the platforms to actually share what they see what they collected and what the understand about the environment to this red arrow distributer like OK and I know like 4 people who work at U.B.S. or Mobile Robots this is this sounds really really trivial.

[00:14:04]
But I'm going to talk about what I'm going to say is this red arrow why I put a red is because this is a very hard research area in the Marine robotics side OK I'm going to tell you why this is a very hard research area OK. So let me give you a little bit mathematics descriptions about this in our confirm in the motto OK so how do we construct this generic environ model for mathematical side is very straightforward you're really when you model spatial data what you do is you find some spatial basis functions OK which is this.

[00:14:37]
Can be any function can be in your network it can be RB as anything like and then the key is you have a set of unknown pair meter saida that you want to estimate from data to match your vision Z. OK but in the ocean and also for other signs we often have to worry about this this noise new this new depends on the space but also has a very important property is their specialty car lated In other words you have to worry about estimating the spatial correlation matrix about this noise and this is not easy to do.

[00:15:15]
So so forth but but you know with all this good big data to see you will be able to do it so therefore the goal is to give an observation you want to estimate data and also want to estimate the covariance matrix tomato spatial data but in the ocean things are not just depend on space it also depends on time so how do we model timing the time variation so one way to do it is to construct a differential equation to model the variation of data over time OK so you've got you've got this dynamical system now and then you combine this then we go system with the observation equation then it becomes a dynamic a field in problem that given the measurement so Grania data streams e K. you want to compute they decay you want to estimate it.

[00:16:04]
For. R. is a position. All right so this is the mathematical framework there's nothing surprising for people who work in Slab this is sort of like bread and butter every day and so so so why where are we wasting our time for so we're wasting our time in actually answering the following 3 challenges right now so the 1st challenge is even though the mathematical model is very very straightforward however we don't know exactly what they are OK.

[00:16:39]
And also in the ocean there is a special challenge localization and tracking and it's very difficult to get accurate localization in the ocean OK Not not like in the air or on the ground and also another big problem with that distributed autonomy architecture is it's very difficult to share information in the ocean the communication is a huge bottleneck OK So so this talk is going to talk about these 3 things that's a long introduction OK so now I'm going to talk about how we can solve these 3 problems and to be honest with you we didn't solve any of these problems but we want to report the progress that we that we have made OK All right so let's look at example this is one of the ocean experiments we did about 2 years ago in collaboration with Grace Reef National Marion sentry this is actually a special spot in the ocean off the coast of Georgia because this is the only place where we have coral reef in Georgia and it's cold water Korea there are a lot of interesting fish species in this area OK so people you're to go diving to Florida right they don't come here so the folks in in this great reef what they're interested in is to count how many fish to be says in this area every year OK So the way they do it is they put this little acoustic tags into the fish either swallowed by the fish or the fish.

[00:18:06]
And then they have this acoustic receivers installed at certain places and trying to listen to those tags and based on the number of different tasks the that they can find and they can decide how many fish to be says how many different fish the seas are in this area for each fish to be says how many fish you can see every year.

[00:18:27]
To be honest with you actually I feel this is a very rough process because for obvious reasons if you put a sensor for a certain area if the fish is out of your detection range it doesn't mean it's not there it just means you didn't see it so but at the current moment that's considered We didn't see it OK So so there are a lot of room for improvement for data collection for this for this problem that's why you know we get our involvement in there so so one of the experiment did is the title calibrate the the tag and the receiver relationship so want to see how far that attack can be detected by a receiver under this ocean conditions so the grease rate of folks they actually install this.

[00:19:15]
Tag and receiver a ray in this piece of ocean so each circle the radius is 200 meters at each location we have a tag at least one track and also one receiver OK in that configuration so the taxi painting and there's a receiver keeps listening so we want to see how many tags can a receiver detect given a duration OK So that's the idea so we want to collect that data so and to explain the results later I want to point out that in this location this arrow points to the Longshore area and this arrow points to the cross for our area again all right so the difficulty about this problem is we actually have no idea how sound propagates in this piece of ocean.

[00:20:06]
So we have no model and we don't know how we just simply don't know the problem getting this piece of ocean what we have are the only what we have are only the counts OK which is the number of detection per hour over numbers pings for our OK And you can see this is the raw data you can see some patterns some variation patterns that some time is larger some time is.

[00:20:30]
Smaller and over this is over like a month but then over within each day there is some kind of also lation patterns so so what does this really mean if you think about it from everybody's point of view this actually means the fish come you get from this type of system is very unreliable.

[00:20:51]
OK so you really need to think about how to improve that accuracy. So we're trying to the 1st thing we do is we're trying to explain variation OK So the hypothesis is we believe probably the tidal currents play a factor in the fact the detection rate so in this area off the coast of Georgia PI's is a dominant current of antis move like an arrow in a circle over a takeaway over over a day so it's keep rotating that that's how tides move and it's confirmed by these real time measurement by a Doppler current profiler you don't know how you don't have to know how it works but all you need to know is that you can actually measure the the strength and direction of the current and you can see that at off the coast of Georgia in this area the along shore flow is larger than the cross or flow.

[00:21:52]
And this is also confirmed by several glider deployments that we did and there is the on the right hand side is the glider estimated slow speed and also confirm that the cross for flow speed is higher than the last or flows of flow speed so we believe that this pattern has some effect in the rate of these actions acoustic rate of detection.

[00:22:15]
And our data analysis actually confirm that we correlated the moment when I say when I have maximum detection rate which is the moment when the tides at a certain direction and we find out that the maximum the moment of the massive of the 2nd rate your only happens when the flow is crossed for the number the cross or flow is larger than the alongshore flow that you are when the maximum detection rate happens but then we want to understand why OK for that we need to understand the physics of some propagation so this is.

[00:22:50]
A simulation of sound propagation in one specific piece of ocean you can see it's pretty complicated and that's all you need to know OK And and then you know we goes into the weed and we try to do some data correlation analysis we come up with a explanation so the idea is when the when the flow moves across or it's it's a maximum flow and actually makes the water better it's makes water well mixed so then when the sound propagate because the water is well mix the channel with increased and it allows more sound energy to go from one spot to another OK It's like your communication channel suddenly improved and you can pass more data through this communication channel and that's why more of these attacks can be detected.

[00:23:39]
OK And this is only one factor we can point it affecting the data rate and it tells you some of the challenges that we face about data collection in the ocean is nothing is. Changed in the ocean it's everything is time varying and everything is coupled OK all these different factors are proposed it's really hard to get to write down a very simple and accurate mathematical model of the process of your study.

[00:24:07]
But nevertheless in the future we talk about putting sensors in the ocean we talk about localized assets in the ocean and we really need to and an acoustic is probably the only way that can help us to achieve this goal so we really need to understand how the sound propagated in the ocean we really need to understand the property as a function of both space and time so what can we do well you know this is this is a good area for robotics video students I don't know whether you believe me or not because we can use robots to actually collect data and construct 3 dimensional map or 4 dimensional map about this sound energy distribution I believe this is a probably a feasible way to tackle this challenge for future ocean science so therefore we started doing some research in this area so this is one experiment where we put in a U.V. This is called Eco matter and we just wrap a receiver around it and then we just navigate it in the ocean in the lake in unity Alabama near user Alabama and we're collecting the sound measurement.

[00:25:14]
And also we clever ways people in China where they put acoustic receivers on their gliders OK and then they move the gliders in the ocean and they got sound so because these 2 things move differently so for the U.V. we've got a sound distribution over horizontal plane and for the guider we actually get sound distribution in the vertical plane and then we can use different signal processing techniques to actually interplay this they.

[00:25:39]
This is a max map construction technology if you think of if you want to make comparison with US land with robotics OK Of course there are a lot of room for improvements here. So so this is what we started doing but then we realized that actually if you use only one big hole you're confined by the trajectories of a cult you can only move in this certain patterns to collect the data so why not just use Marvel vehicle at the same time so we do we do swarming we do corporate control in the context of this problem and the idea is to say Well suppose you can do this then you can actually adaptively track a curve constant level set and then and then gradually construct a better map a more accurate map of acoustic distribution in a certain area but then we never got be able to do this in the ocean why well because of all these other practical challenges because of the cost of the experiment because of all these other things and in fact that's one of the things I feel that's hindering the development of Merino body is because if we want to do experiments on the real robots every day here at Georgia Tech where should we go that was the 1st question that C.S. asked me Seth asked me to to you know to answer OK so I have been thinking about this problem for at least 15 years in time when I come to Georgia Tech or I always bothered by this problem my students wanted to experiment every day you know because you guys in the neighborhood doing you know humanoids and all these you know really exciting things and we want to do something like that every day too so this is our own server so we found that actually flying planes has very similar dynamics with underwater robots OK and we can do swarming easily with a blimp we developed this is Georgia Tech autonomy is planned and this is a smaller version nothing is very safe to graduate students if you do this every day you don't get hurt not like some of these car rotors and and all these other things and they fly for a long time they fly for 2 hours per charge not like some of the car olders only fly 10 minutes 15 minutes and you have to charge them your worry about.

[00:27:39]
Flying of swarm in the same time and we integrated with Matlab and I think everybody loves math up here I know some people like Ross and we have a rough interface now we have entered we have the simulator we're going to we can make it open source very soon so the students who are doing it is sitting over there.

[00:27:58]
So and then we can do this experiment every day and we can put up demos for high school students at a mentor schools in K. through 12 important visitor from all in our in as a. Professor or some other universities so anytime you say I want to demo from your lab it's ready OK So this is this is our blend and we found is a very powerful too to support Marine robotics research and a lot of other things human robot interaction with the sensor networks so dimensional slam you know all these things so OK so I'm going to talk about the 2nd time now underwater localization OK so underwater localizer is really hard in the ocean without paying proper Pads So you're only a U.V. do on the water is they can rely on passive beacons you have preinstalled beacons that give you signal and then you you localize yourself sort of like off the track system in our lab or you can use active sonars to actually shine beams onto.

[00:29:00]
To this allow marks and then you get some relative or local additions sort of like the vision based you know Adama tree idea land. But all these things all the sensors you use all the system beacons you use are a least 10 times more expensive then anything you do out of the water.

[00:29:20]
So and another big challenge here is in the ocean localization is very much coupled with the understanding of the ocean flow OK because the ocean constantly move you stand you stand on the beach one of the things you enjoy is to sit there and watch for the oath watch the ocean for 2 hours while you can do that because it's changing and you cannot predict how it changes right so therefore you have to it's a it's a very attractive but on the other hand it's very challenging so you really to have good localization performance we need to really have ocean models that tells us how the motion moves and suppose this is the real ocean pretend and the model we have are actually like this OK the resolution is not very impressive OK And and how does this play into your local ideas localization performance it took was about 4 or 5 years to figure this out again the mid we're not that smart.

[00:30:16]
And and the way we do is we actually do it this way so we kind of simulate the vehicle motion in the ocean model and then we collect the real data from the ocean and we compare the 2 We call this the controller grounding prediction error OK because why we do this because there is something called a grounded prediction error so we just have control to add you got it right so so we want to predict how so what is so the 1st the 1st OK The 1st problem is so is to find this throughout ical bound on this error in other words what is about that you cannot beat and you have to accept when you are guided by a not so good in the low resolution of your model.

[00:31:02]
So you actually boils down to a nonlinear dynamics problem not even control it's not here then I mix problems so even if there are an F. is the difference between the. The model and the real ocean and the challenge here is it is a non-linear model but also it is to cast a so as a random sample from a from from your from your modeling system OK So the idea is where did he go I think a lot of people took 65 to so you know that D.D.T. has at the this is a qualifying exam problem that evil go to one of the criticisms OK and which economies will go to you will go to the stable curve EMS out of the 3 you could have here which one is stable the blue ones are stable.

[00:31:43]
So therefore we're able to use this to say that we can't hack it because everything is still Catholic and random we can calculate the distributions of the stable you curve and the through out equally the distributions like that OK it's not a Gaussian of course and we can drive right on formulas for this OK And the interesting fact here is now we can talk about fundamental limitation because you see where the meaning is the meaning of the position of the secret M. is actually 1.5 times a great size in a one dimensional setting in a 2 dimensional setting is about 225 great size so this is how well you can do if you use a ocean model with a great size H. your local ideation assisted by this model cannot beat 2.25 H. in a 2 dimensional case.

[00:32:34]
And this these are some simulations confirming this the green one is when you only consider this model resolution the blue one as you consider some other effects and orange one if you combine the 2 you can see if you convert to this simple equipment very fast and you stay there and then the other factor kicks in you kind of start slow growing slowly growing but you cannot beat that.

[00:33:01]
And it actually confirmed by real life data that we collected in the ocean experiments so this this experiment happened long time ago this happened in 2008 and we observe this we observe something in there that inspired us to do this research and after about 4 or 5 years we figured this out.

[00:33:19]
And you know at that time we actually cornered 6 gliders in the ocean and this is actually started the research on mobile sensor networks is very fond of the work. And what we observe here is you know we were using 3 different ocean models of completely different ocean mottos but the only thing that in common is they have the same resolution all right and to guide of a cause and here's here are the arrow girls of the control a ground in prediction error and you see the patterns are very similar so they also low down at this so there is a fast growth and they also slow down at this moment which is about $2.00 their great size OK so we've served this and then we try to explain this and we found that that's a general principle OK And this is only a very small progress forward.

[00:34:05]
And also you know in our lab if we want to study ocean current what we do what we do is we use a centigrade winds which is ocean current and then we use the blimp to map it to try to map this wind field pretending we're mapping the ocean and so we deploy the blimp which is the motions effect by wind and I like the car rotors so it's perfect by when and from the trajectory we can actually do a learning algorithm to actually you know try to learn the windshield Yeah and if you only have one trajectory given to you going to learn it's going to be wrong why.

[00:34:42]
Because the process in the citation condition in adaptive control. So therefore you do a lot of this experiment a lot of trajectories and you get a lot of estimation if you use them together. And this is an area where we have pretty good data and this is the reconstructed windspeed map so where is the fence.

[00:35:02]
On the right so it's a you know it's very obvious because the jet flow All right. And different from the ocean experiments where oftentimes you don't have the real ground truth of the ocean flow data in the lab we can get the ground truth so this is a mobile robot we we you know modify it on the direction of a robot and we put when sensors on it and we do this automatic mapping of Winfield using this mobile robot sensor autonomous Ali You just hit a button and you said their weight and then this is the code I'm told ground troops that we have observed is actually now the ground troops do to the difficulty of marrying low speed wind and that's still an open problem that we don't know how to do it more accurately but this is this is reasonably well it shows you the pattern the Jets low pattern OK.

[00:35:56]
All right but this is all these experiments are carried supported by off the track system there are localization in the ocean you don't have a localization all the time unless you go to the surface from time to time and getting the G.P.S. measurement so so often times here's a situation we face so so suppose a glider is on the surface of this ball point and then we predict it's going to surface that they're OK and then we'll let it die and then after about $56.00 hours fears over there and there's a big difference between the predicted position and the real position and we call this motion integration error so suppose you only have this information.

[00:36:35]
The starting position and position how do you find out the trajectory in between and how do you estimate the ocean flow in between the 2 positions OK and that's a very interesting question so. All right so you know and it's also a chicken egg problem because in order to predict the motion you need to fill them in the slow you need to know the motion so why don't we just so with iteratively in fact this is exactly what people do when you go to a hospital and get a C.T. scan so the idea of the city scan is you have a single ammeter that shines some rays through your body and A at the end there are.

[00:37:19]
A ray of receivers that get this signal OK and from the difference in the signal strengths you can run a mathematical algorithm that solves the inverse problem and then reconstruct the city structure of the organs inside a human body so we can borrow this idea so here's here's what what what we're going to do is 1st we base our current knowledge of the ocean we predict the trajectory of a culture to actually adjust the prediction so that we can reduce the motion integration error and then based on this new trajectory we're going to update our estimate of the flow and and try to again re compute the trajectory until the motion integration error converges to 0 so this is what we call motion tomography OK motion tomography I believe this is a pretty specific thing for the former body aches you know I often have people ask me So what's so special about marine robotics is there anything special here you know here here's something pretty special that normally you probably don't see this in the.

[00:38:25]
Ground robot or air or robot. OK So let's compare basic C.T. and P. So in C.T. you have known in your sense a race in Mt that you have a gnome trajectory running your trajectories on the water very close in C.T. you're leveraging the signal attenuation emotion tomography you have motion integration errors and you see if you use your used car smash mouth of the silver then your set of different.

[00:38:51]
In your set of algebraic equations on the empty side the problem becomes non-linear OK So therefore the algorithm on and none of your side is called nonlinear cosmos Newton algorithm and I have good students here trying to prove the convergence of it for different versions of it it turns out that that's not very trivial the convergence but of course if you run it it most of time is converted works.

[00:39:14]
OK So back to experiment can be verified as them in our experiment using our plan we can't so this is to the appointment one overlay together one point with the wind another deployment without the wind and we're only using the data of the starting position and ending position OK we're collecting that starting position data and then put this in there even though we can see the trajectory OK but we pretend we don't know the trajectory OK after this experiment is done.

[00:39:43]
Here are the flow field we constructed using motion tomography and you can see it sort of have this nice nice jet structures OK and it shows you how it works and also we were pretty lucky that. During one of the previous experiments we have on gliders we have one glider that started wandering around in this region in a small area for about 3 or 4 days and with this you can think of this as a glider repeatedly scan a place for 3 or 4 days and allow us to apply this motion tomography algorithm to this real glider data collected in the ocean and on the right is the flow that we constructed actually physically make sense even though we don't have ground truth data to actually physically make sense because it's kind of close to the jet Gulfstream jet stream so so you can see a jet stream there OK.

[00:40:36]
All right the last one is information sharing difficulty what we have been able to do now is we have been able to come up with a index that measures the quality of communication and quality and we call this requiring quality OK It depends a lot of things suppose you want to transmit M.P. biz of data and the package sizes and the using using this link quality you can say OK this is how good I want to channel to be and this in that actually depends on the bit error raised the time to transmit one bit time delay and also some scaling weights How do you obtain this again you have to collect data to obtain these Yeah so everything is coupled remember I said All right so if you know this required in quality so suppose you have a channel with a bounded thing quality what we can do is to use this inquiring quality to actually calculate how many bits I can send reliable per packet.

[00:41:38]
And with that knowledge we can decide if we can do we can apply all these great image processing or deploying algorithm to actually parse the flow field or any field in the ocean into features and we actually can extract the number of features that are compatible with the channel and decide whether we're going to transmit this feature or not OK So on the right hand side you see some of the representations of compressed representations of the flow field on a lot and side OK and depends on the number of bits you can transmitter about reliability you can mix you can.

[00:42:17]
You can have different details of the boundaries of the feature regions OK. And also another interesting thing we discovered is were distributed past planning if you're sharing this feature only maps not just seeing the whole maps you can achieve in your optimal path planning performance in a distributed way OK when I think your ultimate Ultimo is when you do pass away in a centralized way where every robot knows everything the entire information the flow but only using the feature maps you can actually have to pass it's very close to the rail pass OK with a lot of information removed.

[00:43:01]
And this is this no 2 ways information sharing but like I said information is very difficult why the difficult can be verified can I show you examples so recently there are trends of developing really small underwater because this is my core U.V.'s and on the market you can you can buy some of these vehicles like roofing robotics century riptide micro U.V. or Munson This is from Germany and here at Georgia Tech you know actually these are these robots are not that difficult to build if you don't want them to go to the ocean.

[00:43:39]
As sure as hell we don't have ocean here so we build our own version of the micro. Georgia Tech major underwater robot. It is actually open sources you can download it from it. And it's pretty useful for us and then one of the things we did is we put another.

[00:43:59]
Like available acoustic modem at affordable price made by where China so on to our vehicle OK and we're trying to evaluate communication performance not to the ocean not in a ocean but in a swimming pool a lot of people laugh at me when I say this the hardcore Marine robot is laugh at me when I say this because they say well if you're not going to the ocean you're just faking it I'm going to tell you I'm not thinking it because I'm going to tell you that the experiment we did in the swimming pool are much more difficult than the experiment in the ocean and you will see why.

[00:44:34]
So this is how we put the acoustic modem on to our robot and we have this weird shaped swimming pool one of the apartments here I don't know where my students know and. And here's how we communicate on the water so we have a clip the modem we shine the.

[00:44:55]
Sound beam to the bottom of the swimming pool and it bounce back and forth and then get to the receiver and the receiver get the message OK So a lot of people ask me why don't you point the transducer directly at that receiver because that's supposed to be how communications are be established on the ground but the problem here is we want only directional capability we don't want to control the robot to always face there we want to robot to move around OK So shooting at the bottom is a good way to achieve that so I don't have to design sophisticated controllers.

[00:45:31]
And here are the transmissions has. Done in a swimming pool and you can see the red means bat so it's. Certain spot that the performance of pretty bad so we have this package operate high and be there are 8 high a certain spot of screen for our modem was designed for ocean for communication the ocean beyond one kilometers is very good model but in a swimming pool he has this problem.

[00:45:58]
And then we go to a lake OK in the lake we have a similar environment we put our vehicle in the lake OK and we found that you know in a swimming pool what we see for you level meters. And only happens in the lake when the distance between the transmitter and receiver exceeds 100 meters.

[00:46:18]
The See there is there's a value of working in the swimming pool right. So why is that so so we work with this modem developer we show this and we're trying to come up with explanations Here's what we see this is a comparison if you're a digital signal processing guy you know what synchronization signals mean if you're you don't understand it's OK I need to know is that they screw things up OK So so in this acoustic modem there are these.

[00:46:46]
Sequence of pulses used for synchronization so in the lake and the pool you can compare in the pool this think when addition pulses are smeared more severely than in the lake why because even pool has a lot of walls OK Many Marty passes fact many of these things very challenging and more challenging than in the open waters.

[00:47:09]
And on the right hand side shows you the waveform of a single pulse and you can clearly see that in the like the pulse of much better in the polls you have a clear winning post in the in the poll was sort of that the winning PAS is not that clear so that that that's causing the problem of course this information we feed back this information to modem developers and he will do something with it he's actually already done something with it I just I just don't have the results that show OK.

[00:47:36]
And this is only transmit when the vehicle doesn't move what we really want is when the vehicle. Moving it can also talk it's already achieved for Dayton after 444 hour phone calls on land that you can drive and you can call but I don't water this has. It's far from being solved OK So therefore it's very important we do this moving transmission paths.

[00:48:01]
And what we found is you know suppose the receiver is at the bottom and you can move across or you can move back and forth towards or away from the receiver because what the performance is the worse when you move toward the receiver. At a constant speed you know why that's so again it's very interesting to think about it so if you think about the echoes around you're running and you're you're missing echoes like a bat so so you know when you are when you are moving to all the target the you know the the paths the you know the change of the channel property actually is much more severe than when you're going away all across so that's actually the hand-waving way to explain it OK Again you see you know doing many robotics in swimming pool in the air is worse well.

[00:48:51]
I don't know OK so let me come through the talk so one that said I want you to take away was is in the Marine robotics if you think about how many problems the central question is how the Enable control of running data streams. And also the current research is turning to was distributive autonomy architecture where information sharing are.

[00:49:13]
Geared to Wald different agents and we're trying to enable that red arrow mathematics is a pretty straightforward is it is a very challenging filtering problem but the key here is not that that that that sort of slam type mathematics the key here is how we can have better forms of models how we can have better localization and tracking and also how we can have better communication performance and I want to thank everybody who worked with me on this topic some of them if you don't appear in my slides.

[00:49:45]
I can invite you for dinner don't don't worry so and I have worked with a lot of people gee you know you can you can tell like so many things in this in this talk and again apologizes if you felt like there is a topic that I know less than you that's completely normal OK Thank you.

[00:50:09]
Well. Yeah I actually there are there are there are terms there that that that we do. So we have after 2 versions of motions longer you have a parametrized version and parametrized version so the parametrized version we try to reduce the number of unknowns that we want to solve because the answer metered version is under determined problem and you need to regularize that so yeah that that we have 2 versions of it yeah.

[00:51:02]
Thank you thank.