SoundCloud Widget Placeholderhttps://soundcloud.com/uclsound/ep-3-self-driving-from-big-picture-to-in...
Episode 3 - Self-Driving: From Big Picture to Individual Preference
Hello and welcome to IM @ UCL: The Podcast, a podcast about the research at UCL that will revolutionise the future of driving. My name is Cassidy Martin and I am your host on this journey of self-driving discovery.
When it comes to reaching aspired outcomes, it's all about planning. For instance, let's say you're an undergraduate student who has aspirations of getting into an elite postgraduate programme. Well, in order to give yourself the best chance of getting in, you will need to plan out your time in a way to make sure that you can check off as many boxes as possible of what admissions wants. This will include setting specific goals with well thought out step by step tactics of how to meet them. And this same idea of knowing what is needed and specifying goals and tactics can apply to aspired outcomes in any area, including vehicular performance.
For this month's episode, I spoke with two academics who are utilising very different tactics to reach their desired outcomes. Our first guest is looking at using simulations of large-scale maritime operations to get stakeholders on board with proposed changes. And the next, is looking at using maths equations to teach vehicles what each drivers’ navigation preferences are.
Just a note before we start, this podcast was recorded remotely. And sometimes you can hear a bit of background noise. Hopefully, this will not distract from the insightful content from our guests.
Let's get started.
For episodes one and two, we mainly focused on how the IM @ UCL facility could benefit from staff research. But for our first guest, in addition to their staff research benefit, you'll also get to see how students can benefit from the use of the facility as well and possibly even give back to the local community in the process. But before we get into all that we’re first going to learn how his own research will be benefiting from the use of the facility.
My name is Kamal Achuthan, I'm a lecturer at the Centre for Transport Studies, part of the Civil Engineering Department at UCL. My research is into decision support systems, the current focus has been on maritime and supply chain and logistics related things, particularly on the risk management perspective of it.
Although the logistics of goods coming in and out of ports might not be something most of us think about regularly, it's actually a vital part of how our society operates.
So it happens a lot behind the scenes. But we all strongly rely upon it, particularly in this country, for all the goods to be varieties. And it’s an island nation, and every country goods that come to this country has to come by sea or through ports. And almost 95%, except for like 5% that comes through the Channel Tunnel trains and some through the flight, but 95% comes through the sea. And so, it's all handled by ports in the UK, which are very, very critical with such a strong reliance on imports compared to exports. So that operation has to be smooth. So, what we are seeing over the last 10 years is quite a lot of disruptions or scenarios that could significantly disrupt this movement. And the country as a whole was dependent on like 10 ports. So how do we keep these 10 ports running without disruption? And these boats are so big that they cannot be replaced in one go.
So when 95% of commercial goods are moved through seaports, and 10 ports handle the majority of these goods, their functionality is absolutely crucial. Because if certain goods don't come in, it could be dire for the country.
The supply chain industry that is more critical, are the ones that are the odd chemicals, the pharmaceuticals, which we don't really associate to be critical, but ended up like the carbon dioxide shortage that they couldn't use in any of the catering industry. But nobody thought that is an important commodity. We are even finding new commodities every three, four years that becomes critical that we didn't know in the first place that it's critical.
In case you were wondering about the CO2 shortage, it happened last year. CO2 which is used in farming, carbonated drinks, alcoholic beverages, it promotes the growth of plants produced in greenhouses, it's used to extend the shelf life of packaged food, and so much more. And when the two main places in the UK that produced 60% of the UK’s CO2 supply stopped working, all of a sudden there was a huge shortage and a desperate need to make that supply by importing from offshore. So, this is an example of the type of problem that can come up and potentially have reverberating effects on everyone in the UK. Now at this point, you might be wondering how Kamal’s research in maritime logistics problem solving can benefit from IM @ UCL’s facility, which is mainly focused on automotive vehicles. Well one way, is that his research team could programme the simulator to act like a ship.
It will be a schematic visualisation of what happens in the port environment, right from the ship entering the anchor point, all the way up to the port environment and beyond rails carrying to the endpoints of, let's assume, energy supply chain of whatever it is. So, you will be able to see the whole movement of what's happening on a day to day basis. And then you will have a lot of sliders and tools where they can break, adjust in terms of resources, etc. Assuming a scenario of it could be COVID, it could be some simple accident that causes certain functionality to be lost, etc. And then they can see what happens in terms of delays and accumulation of goods here and there and which resource has to be completely switched off, etc. So, it is not exactly one view but all stakeholders looking at the entire system in a single screen.
And creating visualisations for stakeholders as well is really important.
When it comes to risk management, they need to have a slighter wider system perspective in terms of what is coming into them, and what is going out of them, and how things are happening. So that visibility was missing. By creating a system like that; and involving them to open up them to see how things actually on the wider scale functions; and also allowing them to see how important they are in that big chain of events; and what they could do to actually bring about this whole system more sustainable and reliable, resilient; is what interests me. Because that understanding doesn't happen if you go and suggest them something about their specific operation. By opening up them to the whole system, and what happens really to the end user, and also how much the government is reliant on all this for the UK wide things, all these help them to understand and probably invest them in the right places to improve the resilience of the systems.
So offering this wider system perspective could really help to get stakeholders on board when Kamal comes up with solutions to potential problems. But this isn't the only way that Kemal could benefit from the use of IM @ UCL’s facility.
So the teaching that I do particularly interests me a lot is the one that I do for the first year students of civil engineering, where we run scenario based teaching. It's a one-week intensive problem-based learning where the students get exposed to real world problems, and how they could think of a solution either by applying what they've learned in that first year or opening them up to all sorts of new problems. And they need to think about that over the next three years, that they can focus on different topics that would help them relate to solving these problems. What we specifically asked them to do is go and visit a big area of streets, essentially Camden at the moment. Camden as the case study area. And so, they do join us a group and then go visit the streets, try and understand what the problem is when it comes to safety, sustainability, and particularly environment-related issues happening in that street. And then they come back and brainstorm different ideas as to how to improve those street segments that they have been allocated by different design solutions. Particularly being first year’s, we tried to be asking them to be as open as possible and not limited by practicalities and so on. One piece of thing is actually they go and try and understand how people behave in the streets and what problems they face. So they literally observe for two days – trying to understand what their streets are made up of, why people are behaving certain ways, what are the missing facilities, what could improve the safety of that particular street, what could improve the environment impact in that area by changing certain things – so they observe all this, but mostly from an observer and also possibly as a pedestrian walking along the streets. And also, being first year students, not many have experience of driving or they have a licence to drive. So, their perspective of what car users, and whether it's a vital part of things, is not very clear for them. So, their solutions are very much restricted to facilitating pedestrians and cyclists, which is kind of the goal as well as re-shifting towards sustainable modes of transport, but it's always good that they also understand the need for car for certain uses who cannot use other modes of transport and also for emergency movements of vehicles and also goods movement. So, these have to be in place and trying to understand their perspective would be very, very beneficial. Which I never managed to do just by the street environment. So what I was thinking about this IM facility was that if actually a scenario could be incorporated in the IM facility where they could actually navigate through their streets, using the car driving simulator, they could also see the perspective from a driver and what they see what they don't see in the intersections etc. So, make sure that the car driver also is included in the facility because the students will also get to see some cutting-edge technology of where these future cab-based solutions also. So, there is going to be some awareness of what's going to come in the future. All those design solutions, whatever they put in forward should not be just for the next five years, but also think about all these future technologies that come in and what these drivers will be experiencing would be a nice fit for them to be aware of during the first year of engineering.
He is also hoping to make this student project more local to where IM @ UCL is, at PEARL and Dagenham.
In two years’ time, I think we may move from Camden to where PEARL is based, which is Dagenham, because that proves to be more interesting for us. And so, it may be the facility may be very next to each other. And so, there's lots of synergies for bringing in other stakeholders. And the students should be more excited to explore an area that has a lot more opportunities for them to suggest rather than Camden, which is actually keeping up with the needs of the future. So, I would rate Camden as more up to date in terms of what they can do to keep sustainable modes promoted and also safer. Whereas I think Dagenham, we can do a lot more.
Yeah, I think that's also great for like when you're doing research of, especially if you're out and about and you have this idea or you're thinking about something from like, then you're like, I wonder what this particular issue is, like. I wonder how it'd be with like, from a driver's perspective. And you could go immediately up to the visibility and possibly try to create something where you could see it or actually apply things maybe right away. And then also just, if there is any issues with the scenario, like maybe things are missing or something, then you can recognise that right away if you're constantly looking at it. I don't know. And it's also easier because like you can spend more time doing both because they're right next to each other. I don’t know.
No, no, I agree. Because I think that would be the handy thing because at the moment the reason for being Camden, and I think in the past it was Bloomsbury which is very next to UCL, we always been it's more about the practicality of students actually going and visiting more often whenever they have doubts about what they've visited. And then after doing some analysis, going back and seeing whether it's feasible. All this could be helped with PEARL being at Dagenham, and also it’s a different neighbourhood where it's a lot more car centric compared to like Camden; so it is still in the kind of older ways, a bit more suburban, with a lot more public transport facilities. So, it would be interesting for students to see how they can promote sustainability in that part of the world compared to Camden. So, it's a lot more options, but at the same time with PEARL being there, they can make use of all these additional facilities that will enhance their experience.
By moving first year students courses to IM @ UCL’s facility at PEARL in Dagenham, students will have the opportunity to learn how drivers experience the city and create urban designs that have the potential to actually be put in use in the local community, an exciting possibility especially for a first year student. Now after spending some time learning about the big picture potential of IM @ UCL’s facility, we're going to switch gears and find out about some specialist design potential for the individual.
My name is Laura Toni. I'm an associate professor at triple E department in UCL, so triple E is Electronic and Electrical Engineering Department. And my research is focused on artificial intelligence at large. More specifically, I really work in frameworks on machine learning that are applied to personalised services and decision making.
And the machine learns through detection and recommendation, and this process is called reinforcement learning. A common robot that utilises this is…
… the Roomba, or the robot that cleans the house. That in part is also reinforcement learning especially the beginning because it's put it in an environment that is your home, but the machine doesn't know exactly the plan of the house and nothing so while taking steps; like doing the cleaning as a hoover, right; it also learn information about the environment for example. And that means that the next time it will have a path that is more optimal and the time after again, again, it will learn where obstacles are so it would learn how to take decisions over time, but it's under uncertainty because especially at the beginning, he has no idea on the house plan, for example.
Reinforcement learning has also been applied to a type of system that many of us use on a daily basis, online streaming services.
When you have your iPlayer on BBC App or Netflix App, all this type of app, you do have recommendations that the website or the app provides to you. Behind all that, there is also the aspect of machine learning tools that we work on, let's say. So, in that case, like, it's more the level of personalization that enters more in the project here, but it's the level of personalization that is still related to inferring user attitude and then taking decision on what to recommend based on that inference.
So how would you in, like you have a certain equation for if they're interested in crime documentaries, or something, and then something? And then there's like a different number that's for if they're interested in this actor or whatever?
Yeah, actually, there are many, many ways. Especially in that example that you mentioned, there are many ways. In the others they are more straightforward, but in recommendations there are many ways. One could be, you have a table, right, and you have users in the column. And, and the videos that you want to recommend on the rows, right? And you know some of this. So, you know that some users watch some rows, some content. But you then have some missing values. So, you then try to fill in this table with some form of inference model. So, you try to solve some equations that somehow helped to fill in this table. But then in other ways, also, for example, to say, well, depending on what the user watched in the past, you can build a graph that kind of link users in terms of similarity. So maybe users that share similarity, they will be neighbours in this graph. And then when there is a recommendation to do, I kind of try to follow what the neighbours have done. So, what other people that can be similar have done. Other more theoretical bits are also trying to look at online learning. So yes, there is an equation that says, I tried to estimate what you like. So in the end, they will be like a preference, and they try to estimate. So, then every time I have a new recommendation, I will try to take the video that fits better these preferences of yours. And mathematically it’s simple. Sometimes what is not simple is to derive the equation behind but then once you derive the actual formula, it's not that difficult, per se. Or in the graph is actually more how do you build the graph? What do you assume you know about the users? What can you know more to build a more meaningful graph? Because once you do that, then the rest is very easy. It's just you look at your neighbour, and that's it. So, the most important bit is how do you build the graph? So, depending on the methodology, you have different questions. Some are more methodology, some are more theoretical on, you know, what is the uncertainty of your estimation? And how much can you trust? Or how much I should recommend videos that are random? Because I have no idea what the user like and ask a question about that.
Oh, wow! So all this research, and this use of machine learning and artificial intelligence, so how will you be applying that and this with the automated vehicle?
Yes, yeah. In many inspiring ways. (Both laugh) Let's say, I didn't too much into the details. But one of the perspectives that I like the most of the research I do is to tackle this problem of machine learning and AI, in large scale domains. For example, in traffic networks, or water pipe networks, so networks that are large. And that's because we can use tools from graph signal processing and graph at large that is a way to represent these large networks. So, the aspect of traffic prediction is something that is really interesting to me, because it would be one specific example of how these tools on prediction or large-scale models could help. So can you inverse, for example, the traffic model on a large scale, especially looking at local measurements and get the global behaviour of an entire network, for example, right? So, and then on top of that, of course, the AI videos so how can you then prevent maybe traffic jams or other aspects? So how can you provide a better journey also to the driver or recommend a better journey? So that's something that for sure is in the topic of the project and the personalization also. So, the work that they do around recommendation systems that can be considered as personalised systems. So, what is the product that I should recommend you that maximise the appreciation, right? Or on the movie Netflix or this example. Also, that aspect has an application in the project that is still looking at personalization to the driver, right? So, what is really the preferences of the driver in terms of decisions that you can take in the driving journey. So, these two things together the personalization, and the prediction would be a strong link with the project.
And in order for Laura to test these systems, she first needs a controlled environment.
Then in parallel, I also have research on virtual reality and augmented reality. And that also has an overlap with the project because in virtual reality, especially now I work on how do user behave in this virtual reality settings? And how do user interact with the content? And how we should then adjust to all this interactivity and interactive behaviour. And we work around, can user be profiled? So, do some user have the same way of behaving despite the content? And all these questions. So, also this aspect of not really human machine interaction, but still, you know, human and virtual reality interaction, it would be useful because the project is also related to some aspect of augmented reality. So also, this aspect would fit in the project.
And so what's the difference between virtual reality and augmented reality?
So virtual reality is really when the user finds himself in a fully virtual reality. For example, I'm in another house that is not mine. Augmented reality is an augmentation of the physical reality in which I'm in. So I am in my house, but then I augment that with some extra people partying in the house, for example, that are virtual. And that is a social experience, for example. But I see my house, and just some extra augmentation in there. But that's kind of the difference. While virtual reality is really a virtual world in which you are immersed in.
And so I guess, like I think of the virtual reality, because the car simulators but like the augmented reality, how would that be utilised?
So in this simulator, specifically, I think it would be mainly starting with VR, I agree with virtual reality. Augmented, I guess could be a step further, like when you are more testing in also realistic environments. But then you can have extra information that can still be provided about the drive or few things like that and can still be provided as augmentation. But as for example, for the simulator, as it stands now, I see it more as a virtual reality indeed. But of course, the idea is that when there is one, then then there is also the possibility to have the augmented reality next. So usually, they kind of go along.
So, Laura's plan to start by testing with a driving simulator, and then bringing it to a real environment with augmentations will help her to really understand the driver’s perspective, and what options could be available when a driver is on the car journey.
So, I have one last question for you, which is, what are you most hoping to achieve with this project?
Well, first, be part of a project that could really lead or move a step ahead the field of autonomous vehicles that I think is really important for our society. So, the fact that we are many researchers in that and the fact that we could have a small step maybe, but still a step ahead in the research in this direction. I think it's very inspiring. Then as a second goal, the fact that we could see some of the theoretical analysis that we do, but apply it to a real world problem with real world data with real world scenarios, even if they are virtual but still, it is always the ultimate goal of our research. So, doing that in this field that is of strong and hot research at the moment, I think it's an amazing outcome.
Perhaps sometime down the road, you will get the opportunity to benefit from the personalised navigation system that Laura is working to create. In the meantime, you can be thankful for reinforcement learning currently making your life a little bit better. Whether that's vacuuming your floor, or finding you the perfect movie to relax to on your night in.
Thank you for listening to IM @ UCL: The Podcast. If you would like to learn more about the research at IM @ UCL, you can check out their website at www.ucl-intelligence-mobility.com and/or subscribe wherever you are listening to this podcast so you can be notified when new episodes come out. This episode was produced and hosted by myself, Cassidy Martin, with music from Blue Dot Sessions. It was brought to you by IM @ UCL, which is part of UCL PEARL in Dagenham. And supported by UCL Minds, bringing together UCL knowledge, insights and expertise through events, digital content and activities that are open to everyone. A special thank you to Kamal and Laura this month for sharing their time, knowledge and insight. I hope you enjoyed listening to this podcast and feel like you learned something new, like I have with everyone I've interviewed in this series. Take care, and I'll see you again next month. Same time, same place.