[00:00:05]
>> Hello everybody welcome to our latest stuff in our what is actually going to be a little more traditional We're going to hear about some really cool research that I actually think off campus at Georgia Tech and so our speaker today is very showy Chunda He's currently a research scientist and director of computing here Thursday or in with the new to Nova and his background he did his bachelor's and control engineering at I University and India and a master that University of Florida at University of Connecticut and then moved to do a cooktop or attack and out of the research scientists he studies mainly human robot he and that's what I think is top is going to be of us today and he's one of them pretty cool ortho and teaching added research which is a really ace and back is really the order the outstanding group of scientists and what your search that so he has a pretty solid background bill and the instructional as well as the research he's involved in a pretty cool r.v. program that I got interact with and today is going to tell you a human robot eating I think you've got a pretty interesting title it all from Focus and talent collaboration and before I hop into the topic just wanted to remind you that we all have them in our next week we have a little bit of an optical feel that because of the over the 2 agents the next week for another day off but let's focus on this that one we have before us I think it's all right thank you so much army corps that kind introduction.

[00:01:38]
Let me share my screen. Thank you so much I'm pretty sure it's on the Today I'm going to be talking about moving from her existence a collaboration or 2 it's reliable collaborative robots. So primarily this talk is going to talk about building collaborative robots that can learn to reliably operate in unstructured environments while leveraging and enhancing their interactions with humans so this is a lot of words our trying to break this down and move on to the discussion so I think we will begin with the question of why right and I was thinking about how to motivate a star and I kind of town is Tarek are perhaps the historic context of my development here it also really with the specter of evolution how we transformed into being home as events from the premise of perhaps a less popular depiction of this process of rights there's a lot of call research in after politics in the this and the end that specially that talks about the role of collaboration in our evolution and how it was predominant and one of the predominant features that enable us to be at the top of the food chain like we are being right in fact we also got pretty good at collaborating with other species this is actually an art from a book called in the earth it describes a pretty good story of how to convert it into learn to work with bulls in the audience to be able to defeat me on the thoughts and be able to survive to the talk I'm going to give today is focused on how we can develop such truly collaborative relationships with robots right moving away from traditional notions of that robots just co-exist of the one from the robot it's not damaged human how to move from that to have robots that can truly truly collaborate with themselves and with us that's what.

[00:03:31]
So of course we need to answer this question of who really cares about facts perhaps these are the answers but it is only with the very many applications that this will be used or perhaps because think about manufacturing in violence that the robot can play a significant role in the interim to collaborate with humans and perform the stuff that are not easily automatable right now we have done so so far they can have robots that can collaborative humans so that they can do a part of the time that all humans are not so good at and humans can do this often provides an awesome And of course there is a huge investment in this here and I want to draw your attention to the highlighted numbers on the right it just shows the growth rates you can see how many collaborative robots are projected to be used and bought at impact to the numbers about which sucks but traditionally the 2 of us and for those economists trading even if we are able to completely realize the Congress raising the possibility that it will have to collaborate with other kind of cars and cars are driven by humans in audition that instruments and people on bikes for example and there's a lot of money of course the invested in this area for Hans wiggles and there are thousands out with patients that it makes sense for having robots that can collaborate with humans in order to accomplish a variety of not trying to make it 4 separate human beings and also have more efficient systems.

[00:04:59]
And the next one would be the whole of us to robots are a person of robots but robots and really help the elderly people and people with disabilities to be able to live a more independent life right and this is actually interesting diagram that shows the population distribution in the was asked by Congress right as you can know that a large fraction of the population is starting to occupy more early ages and as this is continuing to grow and it is expected to grow even further as we go Do you really need technology that can Ansari societies question of just being able to support me over and of course there are many notions of how we will be able to think about is robots.

[00:05:45]
And to mention there are people with disabilities who might benefit from having an office a robot that can perform their knots and make them look a more independent life and there are some song that is that just takes in how common disability is and how many disabilities can use something like an office to a robot.

[00:06:08]
So what do we really want from robots in those are being really useful in all the story ideas that we want to show we want to see things from robots right one is a liability the other one is Venice and the other one is quite a mission and to get are we can have robots that can truly collaborate.

[00:06:27]
So how do we really get there what I'm going to talk about today is taking a 3 pronged approach radar 3 different sets of 2 that will help get you one of learning and the other one is estimation back you're able to estimate what the learned model and the other one is struck and you will see an underlying approach across all the different techniques that I'm going to discuss today is the fact that they all have a certain sense of structure so that you don't have to learn Martin from scratch you're not trying to infer things from scratch there's a really strong structure and the structure either comes from Physics are either come from the main not the parts that start a little bit about the library side so the question here that I'm going to put this guy up until I we learn the skills from human rights hoping we learn the skills and the answer I'm going to argue is exploiting structures I mean for amateurs so you're in this particular context what if you want is robotic in working on structured environment and still have to live with performance and be able to adapt to sudden and unexpected changes to be able to leave work in just that look like I promised this is certainly not a picture of my kitchen but just to say the point that we want robots and work in these up and buy and use is an example by on the left you have a robot being taught how to pour water and the Cup So this was actually demonstrated to the robot and then I would know how to pour water into the cup right so what we're really interested in a snag is that the cup might be in a different location that can be unexpected disturbances that can come up but what we want the robot to be able to do that I've got to be changes here in this video I'm actually moving the car and the robot is still in will to complete the.

[00:08:19]
So there's a lot of living going to space if it comes to learning from history you can think about learning directly policies from demonstration you can think about after the vision techniques like imagery important learning that of the idea is to learn the underlying cause function that the not mind and when you think about just learning policies you can think about learning I have dynamical systems ask in Japan and other cities or you can also think about learning the probabilistic distribution of the geometry bipeds back from that learning from learning with primitives and with the learning from that of of that that you can think of our learning stabilizing function that you have that I'm something you learn from the area but it's not inherently stable so you have to learn supporting functions that like the systems I think in energy design stable dynamic of that I've got to be to be civil and be able to talk a little bit about this and even the thing that there are things that I do use the up and out of it methods our compassion his methods in a little bit I'll give you more context with respect to what you.

[00:09:19]
Just situate the related work these are the kind of work I'm going to talk about all over there are numerous trade offs in these decisions right and this isn't a perfect time for music if you wish it was like one of our recent the biggest in the area of l.s.d. and that we discussed several categorisation that the free trade up involved it to see between the skeptically.

[00:09:43]
So I'm going to talk about one of my bruises Clark in fact in dynamical systems perhaps So in this particular approach there are 3 distinct possibilities with one is that this is a method that can turn dynamics aboard their motives right course of a whole and we want to look nice from demonstrations using the traction a lot and the 2nd doesn't vision is that big in getting the convergence of the robot and the backup is a spectacle of what the initial conditions and the dog wanted that he will be able to see how this enabled him standing in the definition sort of like always are if the couple is moved of the **** the robot is moved it should be able to instigate.

[00:10:26]
So here's a plot diagram overview of this process so you start with a demonstration that you can provide an interest in the book instead of doing the orientation paper you can learn all that next so you will take this data and you're going to learn a new dynamic hundreds of instruments which will become clear in a 2nd and you're going to use the stream models in order to be able to generate trajectory through the robot so that it can reproduce it into not use and be able to have a live performance as an output.

[00:10:55]
So let's talk a little bit about how these animals are in court right to be are going to think about statistical Danimal of the things that can be used to encode the stabbing the primary want to learn to bang that into an offense because they want it to be it depicted as authentic and evil but to change this you know the idea is we want to learn the dynamics of these people and we're going to learn that by approximating and learning that kind of what the scene of the state of it using a caution picture model and of course governments model will have a bunch of I mean does this mean the vision of each of your government and I give you this to use as little of the learned model in essence what you would be computing is a cloak on expression for the conditional expression of the event which will give you an idea given a particular state what could be the state of it which means for the robot it been given a particular state how would you move from that right of course you will you have a book on expression which are sort of happened one structure if you take a close look at it it looks like a convict combination of lenient events but of course they in fact are is a nonunion element and so was this contraction analysis that I talked about right so conduction analysis is nothing but differently than as a private he asked the question of how the solution to the banning of the sims it was with a spectator right so.

[00:12:18]
You have to decrease as your incorrect they are nothing but 2 different solutions of the dynamic of them just that they started slightly different initial conditions and I've kind of also watched the trajectory get and let's assume we can think about words with between right sort of distance between the 2 parties and now you can imagine asking the question of how does the different was on time and that's precisely the question contraction analysis and it turns out in front of us inciting the trajectory bukan good use of the meaning the words or the statement will diminish as kind of what's so I mean I'm going to be just lower our systems that are contracting not quite this is going to want to consider the diagram on the left what that shows is that their trajectory is like that you want the robot to be able to read it right and you want to learn the dynamics if you notice the plot in the middle what it's showing is the Euclidean squired distance to the target you can see that it's not monotonic right if you do traditional a contraction out of you will conclude that the system is not contracted in the because of course it goes beyond the target 1st and then goes into the dark and readers and convergence into the dark with all the truth so this means that just taking you being in seconds of distance really limits the kind of distance you can learn under this path so what you want to do is have a sense of generalized force wives actually if you start out if you find a profit metric that defines the distance that is the plot you are seeing on the right you can see that that distance is actually monotonic p.d.p. and you can analyze the system using the generalized quite distance to group contraction so we are going to look at what of general a contraction that the distance measure is now given.

[00:14:12]
By was a documentary that defined a non-Euclidean version of the decision and you're going to think about the rate of change of do this right if you actually compute that in exchange of the good that it will get any question that's in the middle what I want you to focus on is the fact that there are these isolated elements that are dependent on the dynamic that's up meaning if you're articulating the dynamics with caution which is a question mark all those that mark that I mean this goes just because models will appear to secret the one thing once you're out of the equation is that we want to make sure that this is the age of change history all that means is that it will convert those who don't have that good will be monotonic we can ensure that the system is going back so we can design a law on constraint that even enforced you can make sure the statistical having to pursue its contract.

[00:15:03]
In order to not lose because the ad constrain optimization problem what you have 0 is nothing by that a patent is worth the garbage model and the set about of it as well as the metric that you want to look at this is because you thought you might not necessarily know what should be that metric what should be the definition of a distance so the idea is to be able to learn it directly from a tree and the questions you saw in the previous slide are nothing but the cost of nothing but any construction that are and then spends are the ones that were good I'm using traction and houses and the actual content you see are making sure our home by police to a specific point as well as making sure you are learning about it because it was just probably have to have to want to disappear and that we've been able to prove that this that if you're able to learn a Belgian machine model under the specific constraints it ensures that there's a certain that I mean those are 2 contracts right and what does this really imply doesn't last that the robot notion that indeed the converse to the code but there is a slight issue you write so that issue is the cause if you look under the.

[00:16:10]
Under the good these constraints have this for all x. or the what that really means is this is a set up in finite constraints if I come to your straight I mean if I don't work that's not competition intractable parts to implement to begin with a solution that focuses on some of twice the competition so that he can tell he's good at that and constraints into constraints that are not state dependent of course they're going to be a little more.

[00:16:36]
Perfect but what it lets you do is it lets you do need to manage their constraints into a fine it's a constraint there's more competition tractable and you're able to log because so of course if you might be wondering all of that is great by that is actually what So one of the themes you are is a quality of the position of all of the modern people to learn about right of economics right so this is a popular dataset used in the feel that there are several who directed motions that we're going to are and you can see that it's obvious to be able to reproduce many of the trajectory from the data set so it means that if they were to catch up and wind off that happening and not that for all these different dynamics it's the same approach that learning is that.

[00:17:25]
And even more specifically if you notice a lot in the middle you can see that you can learn different local dynamics depending on Baptist this is the advantage of casting this at the problem of finding the end dynamics meaning depending on which a local region of the state that you're then you will be able to learn different times and that can become very helpful.

[00:17:46]
Right in terms of numbers in terms of wanting to be very able to compare this approach with the what I do in of the Arctic needs I think they want to see that when it comes to activity producing demonstration what is the on the left of the particular metric the use that.

[00:18:03]
The adult or the reconstruction had on the reproduction at all we've been what they know what they will do that is what was demonstrated and we're able to see that would be they do it would be to constantly back off then not try and work and that are 2 of the of course that I looked at here are the primary differences the class of metrics that defined and I'm going to be in this right without going into much detail it's just that reference to different versions one with a more complicated notion of distance and the other and of course the most up again of your notion of distance the better you would be able to do it also happening let's that I mean.

[00:18:42]
Ok so the reason why I'm flipping through some of the slides I apologize is because I'm not able to get rid of the high lights I meant to hide in that provision or to better reflect. The other thing is also being lose our orientation time so in addition to position that it's like I talk like you can also learn all the addition and then come to a dynamic the be again trying to combat the approach with some of the approaches that it will to the news organization that makes and the best of the bottom and in terms of this interaction at all except of course now we're in that we're doing it so there's a lot we're doing that are for measuring the difference between what the robots did that's what the demonstration was and identity able to do consistently well come back to the baseline and keep what your cheer is the support also ensures that the trajectory for auditions that you generate are guaranteed to stay within that we're going to inspect right so it's not that you're going to rely on those are not always vision you can make sure you can learn the statistical anaemic of distance in the absolute nations they start with a lot and you look a lot in isolation that and then do that to make sure you think look on expressions like we're going to record getting the to say the protection things so we also did a couple of tests that we had not and I looked at reliability specifically what it means in addition to being able to see those for exactly what it looks like on the robot you're using that the robot is doing it's already not but then I see and I'm there I'm not letting you complete not by asking always trying to do my keep for telling the robot motion but it constantly come through x. and try to complete the task because it has the time embedded in that mix model for how to move from different parts of the state space.

[00:20:30]
On the right hand side you're seeing it from the same video you saw earlier our assets trying to board up what our eyes which the cup it realizes now the good location is different right so you can simply achieve this by canceling the dynamical and it instantaneously adapt and understand that the point is to go over the water into the cup as opposed to a pita to meit's off the next big in terms of the question of what cactus sticks of the popular bar right so we looked at how to learn to live with skills but of course we met the implicit assumption that the knowing was fair in what kind of characteristics we need a captain was but it might be true that in many cases you're not sure what sort of skills you need to walk outside or walk out the sticks it needs attention to our approach to solving this problem what is the learn language when you see different characteristics and also learn how to balance the staff so that's what led to this approach we got must be quite caught up.

[00:21:33]
In this particular approach. Imagine there are 2 demonstrations of this is just a guy example with 2 different demonstrations as a short shelf and now if you insist that you want your system to just recreate the conditions in publishing Why do you want the robot to d.t.s. reproduction What are you trying to do with trying just average them right this would need to constitute predicted for this and of course one good eye view not as moving it in the right but even then it will try to just do the average which might not be the point of these demonstrations You can also think of other difference of whiteness in which you can learn the same which you can define tangential while that you're trying to and they produce demonstrations in the and into wine and if you try to reproduce them this is how their supply correct data off the approximate human the velocity right artist speed of the robot motion but still it doesn't quite accurately capture all aspects of the demonstration Similarly you can go one step further you can think of now is what right that you are thinking one particle is ripped out and once you have the reproductions Imprimis sure that you are able to get a smooth trajectory because it's able to mimic the other side of the demonstration but then it misses the point about the particular location right so what we proposed is a balanced set of learned distance that can show up the dynamics in the accounts of the distribution of predicting all the different wind of distance and be able to combine them so that you can effectively temperatures right.

[00:23:09]
Don't use a lot diagram of words of how we're going about it so you will have demonstration presumably in the competition space and then you know what is whining transformations and what do you do if you simultaneously learn the density in each of these quantum that distance with respect to time you know I can even time that you should be with respect to these 3 different wives and I would You can define a reproduction causing it to be quiet and what we do is define the cost such that it's made by this means that you don't even want to be in a life that what we're deviating from the demonstrations then the demonstrations then so are why he different right what us this is there's something about that who they are there's something about that provision actually with the fingers of the goat in that tells you how important certain regions of the demonstrations to be because they thought that and in addition to that we also look at that of the impact those that will tell you that later with Lewis all the different cars and we do this by going to be learning each way from different 1st and under really if you combine all of these cars to get up and discover that and think and go offline and then it comes to execution time is going to meet all the conduct optimization possible and you can also make sure you saw some constraints such as if you wanted to go to a particular point or you want to go back and go through a point all of those constraints can be taken into account and you will be able to do particular station.

[00:24:36]
So he does this on different not what you say the years of while you did a summary of that all the autumn sticks that you see that the buy what was the most influential quite an example of that right the f. or Nazi or we have a handwriting thinking pressing and pushing you can see that different Why did systems played an important role when it comes to difference what you've been doing a lot at the vet's office so you have these deadlines back only single quantity were taken into account and he also defended on another weird line that we asked the question of is being wanted.

[00:25:08]
Learn events from demonstrations and one not just by going easy on a novel idea print and uniformly made of one it turns out that what they were $400.00 pounds but it doesn't quite do so well and comes to picking President Bush and be able to see that the other is they leave the fact that we doubt that we learn to live influence of these different quite different spots well across all the time.

[00:25:32]
Arek So let's switch gears a bit so we talk to we talk about how we can to live really Lloyd skilled or robots and how they can execute it with these right let's talk a little bit about this dance of the bandits or getting to know something about the park not.

[00:25:48]
The robot but a human right in the interest of time I'm going to provide small pieces of how we think about this so we are asking this question of how can we learn and anticipate human behavior and the answer to that beginning in 4 and predict behavior based on her while if you look at how we can learn from demonstrations of human traits one way to think about these demonstrations is that the human the beating the robot dog that's another way to think about that is demonstrations are also indicative of how they do not write if you have a particular human in a particular the demonstration that there's something about how that particular human.

[00:26:24]
And you can exploit that we can learn those models and use the the model as predictable and that's what these words do right so here's an example that would be have a maximum likelihood approach that trying to predict the school the human bridge it can work with multiple humans and Battle of what you're seeing is a couple of my colleagues back in Connecticut are trying to reach for different objects placed on the table is able to accurately and quickly in 4 of which object that they're trying to correct so this is done because the have predictive models this is because we know how to reach and there is a model that captures how these defendants of their capture and you can design a maximum like you would approach that would be able to.

[00:27:10]
For the package of course the other best to none of the data from the publicly available data set this one is a particular one from Cornell but there are several reaching the us that people do and they're able to use the smart and another official feature of this particular version is that it also update its model as it sees the human right which means that it's not just that it's relying on whatever to get a 2nd was used to pray it although I've got to ask this model at the human is more so presumably a time to go to get better and better at working with a particular human so after this we have this question of could be do something about having a prime distribution so goes out the previous method does not share videos when you have a lot of objects in the scene but a lot of the next step.

[00:27:58]
And next intuition that be explored here is that it doesn't people look at the object they're trying to reach for Seems obvious but how can they can show up this information at the prior distribution and do this for coming on and that's precisely what this approach does so what is the good stuff that's come to the prior distribution so you can sort of narrow it down to Winstead of I'm just talking like you eat by the human do not get a prize and you can use the hypothesis because they're going to bring it out this particular aspect of a human is already what this lets us do is there now it was a lovely thing really flattered by that's right I think if I remember correctly there were about 24 objects in the scene and Other than that it would do really well to Ok the next let's want to see the art of quoting it right so we talked about how robots can learn from humans to reliably off it these are the by our robots can make predictions and understand the human.

[00:28:56]
Operating that we can also think of it's not to go where the robots need one of them saws it up for the Human Being past the office so here are there going to force us this question of how we leverage diversity. When it comes to diversity what I mean by cure I mean but everybody here is under a team of robots right on to human robot that you have to open human sense of a robot each of them having adding their own set of capabilities there are different types of robots the one we are interested in is in its widest diversity of to be able to solve the conflict so the answer that I propose is paid based on finding which means that we are going to look at what are the characteristics of each of these agents but in the scheme and effectively exploit their little bandages and disadvantage to us another 5 months you know what not and I'm going to ask the field of course it's pretty popular that there is a huge set of a proof in the literature are that great to have this problem and here's why one popular way to categorize these approaches right to make any candidate to come along to the divergent one I mention is the task by which means by then you want you need a single robot to achieve the start of whether you need the collaboration of multiple robots and the other one is the robot type between candies or what simultaneously do want to bring us up to do they have to focus on one thought at a time and then there's also this notion of allocation that we start about whether you want to instantaneously assigned tasks to the robots right now or you want to think about scheduling problems that you want to worry about what will work so the discussion today Bill be limited to a single come through what I'm wanting to wipe off and instantly his assignment Let's talk a little bit about the solder them they have been calling strata and this is the hatred you seem I was describing as an illustration of what you're saying different types of robots you're different types of human you lose your job and that's it they all have different characteristics what the other are you are nothing but distributions of their capabilities or traits and they're calling that the straight distribution.

[00:31:16]
And he also can think about not having the requirements and constraints right so that are less important c.r. and each of these knots have their own required for what it is that you needed in order to successfully actually do the stuff and the stiff drink and also being targeted imagine that the stuff where you have been friends political games and it and it is that in this particular kid there's no way out of a robot are human beings not to have one of the 50 bucks to wear or you can want that using the stuff and the other stuff greatest.

[00:31:51]
So when it comes to the not design a problem we're primarily interested in finding how to distribute and a function to a lot off a law does not get off when you're given certain tough requirements if your door These are the requirements for us how would you effectively make sure our robots end up in particular cost and for what of coalitions that make sense and that are required to achieve that if I live in what I want to strip here this being the do not assume that people what would be the decide distribution of the aid in other words we are not simply interested in just moving the robot on that are interested in figuring out that should robots go into further out of the collaborative with each other out of the humans and robots on in so that they can achieve.

[00:32:37]
And we do this by computing optimal transition what that means is maybe they're all distributed in different at different locations at the beginning even back to us about acquiring and what you know that after a while because we're going to look at how to move them around the basket up such that they will end up in particular places that make sense so that the task requirements and the edge so I have one thing on this is what it looks like on the one hand you have the speed strictest meaning your capabilities of your team on the other hand you have these desired stuff exist which means that you have a ton of concrete wired for.

[00:33:17]
That ethos to begin with in your mission now wasn't just that a big mission is being given to the algorithm that because China what it does is trying to pick it up if not to ship it on the route that 2000 effective I mean that I can see that are different coalitions that end up for me for different dogs that are suitable for that particular task so you know the core concretions work the purported unifying framework the model capabilities in the 1st place and look you model what the capabilities of agents are you will not be able to think about it how to conduct your thought.

[00:33:53]
We also want these capabilities in continuous unstrapped way right especially in combat and prioritize look at one of the instantiations of these meanings here are simply looking at whether or not a particular thing existed in a particular agent we're interested in seeing it can be that look underneath and start out the models tend to make fine differences and at least know about the differences for example of the Big a lot of a certain robot is 5 pounds and another robot Spiritus booking you want to be able to differentiate and we also have an algorithm that does stuff assignment with defective boat one we call a sack matching which means whatever the requirements are you need to meet it perfect and the other one of the minimum matching which isn't usually means that you're not going to be penalized for over shipping you just have to meet up and buy if you go out of it that's why if the objective of the 2nd.

[00:34:49]
So you didn't how to take them all right so it seemed obvious he was made of a finite numbers What was because of species which I think boosts the price of an easy to define my capabilities within the species of birds there may be literally him but yet you know that they all belong to a family or a species and that's how we're going to model I think in the diagram you're seeing that is let's say that as an example our particular team the test to create the tree one in trade and what you're seeing is the distribution of different species and this is what we called it the Street model that plays to figure out and model the Gabler you of the species wanted to know what you're what they were encounters this morning that's where it becomes very good with a number of reasons you have one to plaster the agents into this mix it doesn't matter how many you belong to because the big pond it just becomes a numerical value want to metric that that's not in that image the matrix In other words doesn't give the direction of the prop and we have the potential to modern watch human robots and especially when you have a human image you're likely to have a lot of mission even within a particular eat right and this particular way of modeling takes get aboard the differences between species and buttons and we also make note of the fact that some species accumulative And some reason not to know what we mean by that.

[00:36:20]
You cannot have 2 slow agents and make for a fascinating right but if you have different abilities they could add up because I think so we make the expense to take care of our differences in being similar and not submit it and we also model even distribution was so assume that the robot that distributed across the dot com and the distribution of Agent.

[00:36:45]
Is denoted by x. hill so and you can think of the dynamics of this distribution given by the side that equation of that of this and what this basically looks after what the agents are entering and exiting the top right so this will help you model the physical movements audience on this stuff and you can look at the stuff.

[00:37:09]
That operates that because competition to it that dictates how fast you want the movement to be and you can also set limits on how vast each about each type of a thing can possibly talk to the author of wanting to be able are after all these particular morning choices being made does or he can represent this as a senior system of aggression there you have a left inside that cost free distribution on the right inside you have.

[00:37:40]
It into submission of the down species to submission and this is the strictest version which means that the multiplication of the 2 matrix East that represent that little what's on and what are the capabilities of the robot inherently gives you in this model how different capabilities are distributed across.

[00:38:01]
And we saw this optimization that we said Ok you have to decide set up requirements that I have and yet if the team that you have available always to distribute the aid so be implicitly identify how to distribute it didn't buy computing that there's a key the contrary and I mentioned earlier we can have 2 different cause functions against their.

[00:38:23]
Requirements have to be exactly not or they just need to be satisfied and of course what we want out of this optimization is that we want the agent to Congress 5 and you want to believe that one thing but this did that and we also want minimum the notice that the trades us for gas security right which means that you're only going to get at this use an upgrade.

[00:38:47]
On the stand and you want them to be minimally because otherwise you will not be shot by that particular bit of agent that you had in that instance but satisfy pick one so he explicitly minimized the variance in the op. So we evaluated this. Numerical simulation that we had a task we have to fight this fight species with $200.00 agents with an e.p. and a 100 total of 5 traits cleared bridged committed and we had up the total of 100 independent randomized ones and we come back with a very slight nod off assumes that they can be more or less by many factors which means that this model only consider the existence of a trait meaning it can describe what Katty things as opposed to the specific they are and we measure the balance in terms of or different metrics and they can be classified as all the metrics on the left compute the exact matching it up which means that it just looks at the distance between what with the cheat with what was needed on the right it looks at minimum matching which means that it only looks at how much until it be too hot which means if the requirement was a 100 and you didn't 105 that many people are going to attribute all but if you fell at 95 it will give you 5 correct and on the top you have metrics that the actual distribution of the robot's capabilities are predominantly meaning that is not a lot of the issues that in the species and the other metrics you are is a different condition that be measured what would happen if the distribution of the chip abilities are randomly generated which is much more similar to what would happen if you have a lot of him with in essence what mediator of the 4 different metrics then you are off to my thing for exactly right so let's just talk about what this guy to me.

[00:40:42]
At the top what you're seeing is that this is going to be able to confidently do that job when it comes to time to find great requirements both in terms of meeting it and ask what is making sure you don't overshoot too much when it comes to dominatrix and I did this custody from the be right we've got our illustrations of what happened that I was on at the bottom you're seeing what happened then these are kind of 2 things I've randomly sample from a distribution the fine that you are to the average human to do that is in the big blind and it does so that that on average of course on average because there are higher back that doesn't quite hold but constantly on average of the this is what happens when you do exact matching when you're trying to meet the requirements perfect so now let's take a look at what happens when you try to move that into just I need like right so they got looks like a different you are a better way to make sense of them is to think about them as they come from the left or the not the right to the ground or the left the key takeaway here is that this approach is a little more efficient than the best one what that means is it doesn't all of the reason that whenever there is a certain requirement it next year it doesn't do much more than necessary which means that it's a shame that you're collecting the resources that's why you see that for the be implying there is a very high did something of interest much higher up in terms of exactly on the right insight you're seeing that's similar to the previous set up look up the able to understand you meet the requirements back to.

[00:42:12]
An important fact if you are all that aside from the biggest lie of life the computer based on that I'm stuck on drugs what used to look at how many guns actually gun was out 100 years that is the key takeaway then you do not consider of the nuances of the time to be a straight face understood after city effect it doesn't help that just simply considering that by the existence or nonexistence of fate others are in situations that you simply cannot solve problems because it is not the ones enough to meet the requirements of the problem so it ends up in the optimization has now been able to find solutions on the other hand if you're able to get to get to the top and you're able to come to do if you meet the requirements that we also tested this on a game of capture the flag and simulate a game accounts of the fly Have you want to see this great satisfaction out of this inability to meet requirements doesn't conflict with being level it does so we can define a set of tasks they have different requirements and we have 3 different teams it's that one teams follows the proposed approach and the other team follows the baseline approach on the other team for the sake of sanity the author gets it what happens if you just randomly apply the right you do know me here you can see that the number you are playing against both the baseline approach and the records approach do consistently better than grab it just means that it even be talking about a mere existence or nonexistence of base it's much better than not anything at all right than a check that makes sense and they also compare the proposed approach against and you can see that be able to conduct cities with up is there's been been impeached we should have been thinking which means that again the council consideration of the streets and how to distribute the engines are not is much more beneficial compared to the baseline.

[00:44:06]
So what's next right the look as the lead different goods of how to realize the life of a collaborative robot by the look at reliability we look at being able to understand humans and make predictions about them and the look that how they can collaborate with themselves as with humans so there are several questions that lead to a lot of that are left unanswered in terms of the library it's not yet clear up how much structure is in the bank it's not via How much structure you want to import into the model to learn from scratch it's partly how exactly we can scare the items of all who have the capacity to counter them inspirations in Monday with the context and there's also this question of Can these approaches be extended to contractors manipulation as opposed to the we just going to our board is active.

[00:44:54]
In terms of that is big enough questions related to how to treat humans as more than the not I short a couple of examples of how we can make predictions of how humans work but these doing methods that can effectively take those predictions into account in order to add up to robot behavior that that's not merely consider human past just enough for the.

[00:45:17]
And it does a politician that these questions are indeed not that wasn't right so one of the things that was difficult with the Brit methods press and it was this this is occasional Fox It was the assumed that some magic on the group would produce requirements but of course it's challenging to try to survive the question I'm interested in here is can we learn the stuff that was directly from and that is not a risk of the downside so this took out the city albeit in game of initiatives and in fact Professor d. off the tough themself can't be modeled in the referee work that you can get and the and the things like this is the minimum of a maximum amount of the time willing to take what is the best possible that I can become of it I think there's a for them to do it in that can be done about that and we can also think about a learning stuff that quite addictive for me and I cannot be trying to hide it would what.

[00:46:13]
I would be this if I do not want knowledge of my collaborators mentors and students who constitute significantly to the sum of the results that you saw c.r. and in somebody I want to go to by saying that I argue that there are these 3 different capabilities that the robots must possess in order for them to be able to actively collaborate with us as opposed to merely co-exist thank you for your time.

[00:46:41]
Thank you very starry for that look like I actually think I have been working and I probably. Look at her and. If there's any question that people put them in the q. and a or the chap you know I was by then or if I can relate to and you go into more detail than that.

[00:47:07]
Analysis I'm very. Thick but you know you're good so what we are this is pretty interested in is the following Right so. The problem is not learning statistical dynamics from demonstrations and documents a model has been used for a very long time now I think probably for more than a decade in being able to challenge of the statistics out of it but oftentimes these are cynical damning evidence which me when you think there's over them and you're having robots operate along these dynamics you don't have any gap and they will have to practice they will be attracted to what is equilibrium point and times there will be instability models which means that robots can't move and react with what you're interested in it doesn't mean a set of constraints so that these dynamics are getting deep to be stable of course there are some prior work that is look at how to do this using the on a map and it does talk about finding the young and talented it's difficult and of course it's not a necessary and sufficient condition it's on you just the vision condition and that and I would just give you this nice pair work of convergence and indulge you specifically this is a necessary and sufficient condition for the system to be contracting so that's what we explain right so this is what the computer the specific constraints you're needing both on the bat and we do that the government model so that the law that model it's gotten to be stick with just one i Phone 5 benefit is what you get robustness right over and there was no way it is because it knows how to move around that part of the state from any point in the state that it knows how to converge to because of.

[00:48:49]
Course. I don't care as I'm going to ask what I wrote in spring which boxes up again or the like they get a pretty cool place to have their work I have a look at. It. Right so the idea that it is you have these learned models from human demonstrations so let's see this human show you how they move 5 times you've now got that look at them to get out of that don't capture this movie what is the exploited status you mentioned techniques in a model to predict what will happen in the future and you can do this too it's right there is one of our staff to get back to like you in that approach what specifically happened to this good vision is our power grid of this time that's what guides the dynamic of the system to make it and what you want to do is come up with an expedition out of them in this case it was an expedition acquisition algorithm we did find that was going to be identifiable which family to make the most so it's a mass of them like you that the other one is just a big issue is not exactly found to be to fuse right information that you get from it and you can not again this important time to predict because a fish in that sense.

[00:50:06]
So you're leveraging that model that you are either exactly trying predict the way. Exactly yeah. Or a question since there. Is a bigger priority to play in the future of the tie up a human. That's why. Many people Couric. Ok I'm trying to make sure I understand the question.

[00:50:36]
Is a big priority to try to replace humans in working dots or helping him do their job better and I think that of it I mean there's a very interesting question right and also a nuanced one I think the idea would be to. Of course to be brave I pray to God to make menial tasks that human stock want to do that's an obvious I thought been trying to do for a long time now in our vision that there's of course something to be said of us how robots can play out there to grow up in order to make sure this helps humans do that Offutt supposed to be in use so many of the stuff that I'm working on this actually is honestly it's Tuesday night in which there is a need for human to be just a purpose not that of a life and thing to have the art honestly complete all the stuff but it's basically looked at an obvious and tasks and there's a need for the human to participate and that is the human need then business acumen in the vicinity of the robot in the market although a lot can operate in a way that it helps if humans and be able to truly collaborate at it was just merely existing boots on.

[00:51:42]
Or a question. Or 2nd. That looked like it. I think there is one more sorry to see the question there is yeah yeah I'm reading it right now so I'll read it but I you're out of time now to organise a race like south out of the way that we are and that recreate what you know the I'm not sure if I'm part of putting on things that I I I was I was simply saying that this is what we do generally wife is that all you need to do for that is to count to some underlying dynamics and of course the dynamics now depends on different characteristics meaning they go into politics will change and are they different geometry class the type of logical aspects of those trajectory then yes you certainly can use math groups like this to capture different characteristics and effectively combine them to separate them but that's Ask part of my knowledge course to generate noise.

[00:53:05]
It. Looks like that that I want to pay top to get a thank you very forgiving great. Thank you I hope it is well and I thought it was great but it's a reminder that we will have another and I just want to be honest by profession because this is the will be a group seminar with a bunch of the faculty from.

[00:53:30]
The forty's that are high gear on.