To kick off season two of The Connext Podcast, we interview Dan Gandhi, Director of Autonomous Vehicles at Nextdroid. Find out what it takes to get an autonomous vehicle on the road, how to best combine machine learning with deep learning and AI, and what the technical requirements are for complex autonomous systems.

In Episode 16 of the Connext Podcast

  • [1:48] How would you explain what NextDroid Robotics does?
  • [6:12] The autonomous vehicle industry as it exists now
  • [10:29] Why Connext DDS makes sense for autonomy from a robotics standpoint
  • [15:28] Searching for the right connectivity middleware
  • [22:00] How Connext DDS supports AI applications
  • [23:10] Advantages of combining machine learning and AI using deep learning
  • [28:07] What does the future look like for autonomous vehicles?
  • [43:19] What happens next for NextDroid?

Related Content:

  • [Webinar] Secrets of Autonomous Car Design
  • [Whitepaper] The Secret Sauce of Autonomous Cars
  • [Press Release] NextDroid Selects RTI’s Connectivity Technology for Its State-of-the-Art Autonomous Systems
  • [Case + Code] Accelerate Autonomous Car Development 
  • [Press Release] AUTOSAR Releases Latest Adaptive Platform Adopting the DDS Standard

Podcast Transcription:

Steven Onzo: Hi and welcome to Episode 16 of the Connext Podcast. I'm Steven Onzo, producer of the Connext Podcast, and today I'm pleased to present an interview recorded in December 2017 with Lacey Trebaol and Daniel Gandhi, Director of Autonomous Vehicles at NextDroid. In Part 1 of this two-part series, we'll talk about the challenges of developing an autonomous vehicle for a busy roadway, and learn why NextDroid chose Connext DDS as the connectivity framework for its autonomous systems.

Steven Onzo: Other topics we'll cover include combining traditional machine learning and AI using deep learning, as well as Dan's thoughts on what the future will look like for autonomous vehicles. We hope you enjoy this episode.

Daniel Gandhi: I am the Director of our self-driving car program. Right now, the company is basically focused on maritime and self-driving car, so I lead projects in the self-driving car space.

Lacey Trebaol: Do you like it?

Daniel Gandhi: I do, it's an opportunity to actually contribute to the space, and to take my background and try to apply it to this problem that everybody's trying to work on right now.

Lacey Trebaol: And it's a messy, complex problem.

Daniel Gandhi: Yes.

Lacey Trebaol: Which means it's right up your alley. It's one of your favorite kinds of problems. Alright, so let's talk a little about NextDroid, because if you go online and you Google NextDroid Robotics, you find next to nothing about the company, with the exception of a Gizmodo article and some job postings, which I take to mean that you guys are hiring?

Daniel Gandhi: We are definitely hiring.

Lacey Trebaol: Engineers, it looked like.

Daniel Gandhi: Yes.

Lacey Trebaol: Cool.

Daniel Gandhi: We have an office in Boston, we have an office in Pittsburgh.

[1:48] How would you explain what NextDroid Robotics does?

Lacey Trebaol: Okay, yeah, I think one of the things I saw was for engineers at that one. Cool. So if you had to explain what NextDroid Robotics does, how would you do that?

Daniel Gandhi: So I'll take a step back for a second to sort of set the stage for NextDroid. If you look at computing today, if we're talking 30 years ago, computers were an industry, they were their own thing. Now we don't really talk about it in that sense, 'cause computers have become pervasive, they're just part of every industry. That's happening with robotics, that's happening with automation. It's going to start impacting every industry, and we're already starting to see that.

Daniel Gandhi: When being a large player in any one of these industries, this means that you now have to adapt. You may have established a way of operating, a way of developing and engineering for your domain over the course of a long period of time, and now suddenly your foundation's becoming a little less stable. What we're doing is we're essentially creating partnerships with large organizations to help bring state-of-the-art robotics and automation technology into their sphere.

Daniel Gandhi: We can act as the tip of their spear into this realm. We work with them on what is and isn't possible, what their roadmap should be. We will execute on projects that are the beginnings of work in that space for their next generation work. We'll work with their teams to transition some of that technology back to them and help build up teams in that space, so they can essentially establish themselves in a new area.

Lacey Trebaol: So they can grow organically by leveraging, where you guys are the experts, instead of having to overnight try and become an expert in a field that they weren't in before?

Daniel Gandhi: Exactly. And it comes back to also, if I were trying to build a team, starting to have expertise in a domain to know who to hire.

Lacey Trebaol: Or to know what to do.

Daniel Gandhi: Exactly. And so, we can offload a lot of those tasks.

Lacey Trebaol: Massive risk reduction.

Daniel Gandhi: Right.

Lacey Trebaol: Okay, so you guys work with these larger established companies?

Daniel Gandhi: Yes.

Lacey Trebaol: And you go in there and kind of become, at first, you are their robotics team it sounds like, and then you work with them to transition that back to them, so that successful engagement with these people would mean that at the end they have their team up and running, and they're functioning, they can hire, and they can achieve whatever it is they were trying to achieve with us, and you guys can leave knowing they're good to go.

Daniel Gandhi: Right, and sometimes that doesn't mean leaving, sometimes that means that we continue to stay on the edge of what they're doing, 'cause we can be small and nimble. We can stay up to date on the research, and so I would say our emphasis is also that we're taking state-of-the-art technology and applying it to production deployment. We're talking about applying at scale, we're not talking about some academic project or some science experiment. The work that we do has the intent of being baked into production pipelines for these major organizations.

Daniel Gandhi: So they might give us something that's very early and very immature, and say this is something more research-y, and we'll give them results based on what we are able to do with it. But at the same time, we could have projects that already have dates for when they would like to launch them. And being able to say that you have sufficient domain expertise in these realms to know that a production deployment of something is possible, what the appropriate scope of it should be, so you're not going to create something that's an unsolvable problem at some point.

Daniel Gandhi: I think I should also say that our initial push is to have partners in the maritime industry and in the automotive industry, as being two areas that not only are sort of ripe for innovation, but ones where our founders have deep domain experience. So we're building out in those industries, and then also looking at expanding that to other industries.

Lacey Trebaol: So then, you just mentioned the underwater, so you guys are working with people who build underwater vehicles?

Daniel Gandhi: That's correct.

Lacey Trebaol: And then, on the vehicle side, I guess you're allowed to say that ... What's the standard line?

Daniel Gandhi: We have partnership with one of the top players in the automotive industry.

Lacey Trebaol: Cool, and it's similar to the partnerships that you described before?

Daniel Gandhi: Yes.

Lacey Trebaol: So that type of thing, awesome.

[6:12] The autonomous vehicle industry as it exists now

Daniel Gandhi: If you look at the autonomous vehicle industry as it exists right now, it can be basically broken into two camps. You have companies that are trying to develop fleets, so expensive vehicles that are always going to be owned by a major corporation, that have expensive sensor sets, and are going to have engineers and technicians on hand to keep them operational. And then you have the more traditional retail consumer channel. This is where the traditional OEMs and suppliers have generally lived, and how they moved towards autonomy.

Daniel Gandhi: The breakdown is one where on the retail side, ultimately the technology is going to be deployed at a far wider scale than what you'd have from a fleet perspective. You're going to put it in the hands of consumers that have no education about the technology, and it needs to be serviced and maintained by dealerships that also have limited training in the technology, all while operating with far more exposure, far less controlled situations than you would normally.

Daniel Gandhi: So on the fleet side, you can go and deploy a huge range of expensive, large-scale computation. You could deploy a wide range of sensors that have, I don't wanna say limitless costs, but in comparison to the retail channel, like limitless costs. And you can find where they operate. You can say we're gonna operate in certain circumstances, and the way that all of them have been operating thus far has been with safety drivers. I think one company has tried taking the safety driver out once, and that was extremely recently, and I don't know that anybody's ready to say that we're gonna start having fare-based service without any driver in the driver seat.

Daniel Gandhi: On this side, on the retail side, we're basically just having a steady march toward automation. Any technology that you could imagine being in autonomous vehicles or not, when you talk about scale, the U.S. auto sales market is more than 15 million cars a year. If we say that with the new technology deployment, less than 1% will have new technology. But even then, you could easily have 100,000 units being sold in a year with new technology. Those cars will get driven, conservatively, 10,000 miles in a year, so you're looking at just your first batch having exposure to a billion miles of circumstances that how would you have tested for.

Daniel Gandhi: You say that the average age of a car in the U.S. is now above ten years-

Lacey Trebaol: Is it really?

Daniel Gandhi: I think we're at like 12 years or something like that. So, now we start stacking. We say ten years from now, you're gonna be accruing ten billion miles a year of drive exposure to a technology that's still only making up like half percent, two thirds a percent, of the total sales market. When you look at what anybody has really achieved in terms of AV testing, I think the largest, which is over several years, is in the four million mile range if I'm up to date, so we're talking about several orders of magnitude difference, in terms of what you can test and what you-

Lacey Trebaol: Some of the Waymo stuff?

Daniel Gandhi: Yeah.

Lacey Trebaol: I vaguely remember seeing a number close to that in the last articles that I'd looked at. They're way ahead in terms of the miles, because every project that Alphabet has they can use the stuff for, like we need to survey streets, but we can also strap that sensor on here and take some pictures. I mean, they are optimized to get more data, and that's a hard thing to achieve.

Daniel Gandhi: Yes, and so when you're looking at that retail channel, you have to be trying to stack the deck as much in your favor as possible. Everything needs to be scalable, and you have to have some sense of security and safety ingrained in it that is going to say that when I put this in the wild, when it's completely out of my control, it's going to behave in a predictable manner. It's going to operate with some definable balance of when it's gonna have error and when it's gonna have its failings, and that when it fails it's gonna degrade gracefully in a way that doesn't jeopardize people's lives.

[10:29] Why Connext DDS makes sense for autonomy from a robotics standpoint

Daniel Gandhi: And so, when you have to operate at that kind of scale, you're looking for any sort of technology that you can bake in that will give you a leg up. Our projects are structured such that from the design of algorithms and the system, and the way that the system integrates into the vehicle, are specifically always aware of that challenge. If it means that we confine scope of a system so that it may only operate in certain regimes, but when it operates, it's gonna operate in a way that's not gonna jeopardize people's lives. That has to be a foundational philosophy in the way that you build out a system.

Lacey Trebaol: Yeah, you can't bake that in after the fact.

Daniel Gandhi: Exactly. And so, we still have to deal with a lot of the problems that you'd face in the sort of fleet side, where you have a lot of computation. You're gonna have to have more computation than automotive is used to, to solve these problems. And as you start to do that, and you start to say systems have variance, and software needs to be modular, they have to be able to be reused, you want some middleware that has the flexibility to allow me to have a production system where I can deploy it, and have some confidence that it will meet safety standards and make a safety case for an automotive partner to put it into the field, while also allowing me to develop, allowing me to have flexibility in terms of if I'm scaling a project up, and I add more computation and that computation is in parallel, I want to make sure that I have a way to push computation as needed across different modules.

Daniel Gandhi: And so, having something that is very hard-coated to a very specific platform, a platform which may be changing out from under you as the industry is evolving-

Lacey Trebaol: Not maybe, will.

Daniel Gandhi: Will, it is just gonna set you up for failure. So you could see that a lot of people in the robotics industry, they'll use something like ROS. But ROS, it's designed as a research platform, it's not designed around having a production-scaled deployment of something. And so, from that standpoint DDS made sense. And what's more is when you're looking at selling to consumers and moving through these types of channels, anything that is purely open source, where there's no responsible party, nobody who's going to address issues, nobody's gonna guarantee maintenance, it is just not tractable. So we needed to have a DDS implementation from a company that was going to then be able to back that for the time of deployment, which could be substantial.

Daniel Gandhi: And in terms of just long-term stability of the port of something, typical automotive development cycle is six years.

Lacey Trebaol: It is, from research to ...

Daniel Gandhi: Well not necessarily from research, but if you look at how often do I get a brand new version of a car, it's something like around six years, and somewhere in the middle you'll get like a refresh of some form, where they'll change some of the cosmetics and add a few features and tweak things. So if you imagine that somewhere along that six year period, I took a snapshot and said this is what we're gonna develop to, I'm going to deploy that, and so it's now marching along and it has to go through all of its conclusion of development. It has to go through all of its validation stages and its calibration stages, its production line stand-up, its deployment.

Daniel Gandhi: And then it ends up in a retail channel, and now that car drives around for 12 years, a company will try to hold something as long as they can obviously, they don't wanna reinvent the wheel, so they might keep systems based on that infrastructure for decades. You need to have something that is not gonna be ... I mean ROS is ten years old right now, obviously it started with small beginnings. You need to have something that is going to inspire confidence from an automotive player's perspective, that either it's gonna be supported by its makers, or that somebody's gonna be able to pick it up and own it and keep it alive for a period of time.

Lacey Trebaol: It's an actual robust product-

Daniel Gandhi: Right.

Lacey Trebaol: As opposed to an open source community contributed build of something. It just doesn't have the same level. Or even honestly of the testing and stuff they go through, all the validation things, I mean they're treated very differently.

Daniel Gandhi: I mean, you can work with open source things, it's just a manner of a company might take an open source project and bring it in.

Lacey Trebaol: And bring their own thing, yeah.

Daniel Gandhi: Right, and then they'll like really test and validate that snapshot, and essentially it'll be forked off the rest of the world, and then they'll take ownership of it. But that's its own cost to go and ...

Lacey Trebaol: Maintain and do all that with, yeah.

Daniel Gandhi: Right, and having to have the expertise on hand for that specific aspect all the time.

Lacey Trebaol: Yeah, so much nicer if a company can do it instead.

Daniel Gandhi: Right.

[15:28] Searching for the right connectivity middleware

Lacey Trebaol: Yeah, so you guys realized you needed middleware, and you went off and researched what fits best here?

Daniel Gandhi: Yes.

Lacey Trebaol: So how'd that go?

Daniel Gandhi: Basically, we had the option of building up ourselves in some home-grown fashion, which is kind of where you start, but that didn't really have ... One, being a small company, it's a lot of overhead for you to be developing your own infrastructure. And then, couple that with a pathway to eventually having safety certifications and security certifications, and every other certification that you would need to get this out someplace, it just adds more overhead to getting something into production.

Daniel Gandhi: ROS was something that we had earlier considered, and we sort of threw out for reasons we already discussed. I think LCM was another thing that was in that same vein, which is out of MIT, it's what they used for their infrastructure from when they were doing their first challenge car basically. And again, being an open source project, it also sort of didn't meet that criteria.

Daniel Gandhi: And we were looking for something that we could standardize on across the company, so both in the self-driving side as well as the maritime side, and the maritime side has different challenges in that basically if you have an autonomous vehicle in that space, if it's kind of moving in open water, maybe you're not concerned about it hitting something like you would be with a car, but you are concerned about losing it. These are expensive assets, and having them disappear into an ocean is not really good, or it getting damaged and not being controllable and not being recoverable, are all issues.

Daniel Gandhi: And like I said, being small and trying to standardize something that could support both use cases-

Lacey Trebaol: Where that if you guys had your staff having to come up to speed on, so that you could have people that could cross projects, and that once you put that time investment into learning the technology, that you guys were good to go. That could be hard to find, so you considered Ross, and then-

Daniel Gandhi: It was ROS, LCM-

Lacey Trebaol: LCM, that was it.

Daniel Gandhi: -And just our own home-grown solution.

Lacey Trebaol: And then, how did you find Connext DDS?

Daniel Gandhi: When we were looking at DDS, I remember us like messing around a little bit with Open DDS, just to get a sense of it. I think we looked at who were the biggest players in DDS implementations, and in the OMG consortium, I guess is what it would be, and so RTI was on that list, and I think that's how we established this relationship.

Lacey Trebaol: And that was about a year and a half, two years?

Daniel Gandhi: Close to two years ago at this point, I would say.

Lacey Trebaol: Yeah.

Daniel Gandhi: I think we started talking around then.

Lacey Trebaol: Can we talk about why Connext DDS was important as like an architectural decision for you? Because, I always think of architecture, and my mentor once told me this, he goes, "Architecture in any sense is strategic design, right. It's do all these right things at your most base level so that you are given the opportunity, not the guarantee though, that you can realize that potential, and realize that end state that you had."

Lacey Trebaol: And the end state that you guys are trying to achieve here with these types of systems is not something that you accidentally stumble upon, so I feel like going after architecture in the proper way for you guys, was a must. What did this bring to that for you?

Daniel Gandhi: Our software is structured as lots of basically independent modules that have specific tasks. That way they can be compartmentalized and independently developed and verified and unit tested, and all the things that go along with that. Having those software modules then run a system together required some architecture where they could all communicate. We have computation that spans processing units, so as you start to say I'm adding more units, ideally you'd have some form of seamless communication, whereas software modules move around computing resources. It's not like you have to make a rewrite of how everything works.

Daniel Gandhi: And also, there's a lot of variability in the data that passes between modules, and the constraints that they each have. For example, input sensor data is usually very high bandwidth, everything in the system needs to be low-latency, but if I miss a frame of something, I usually just want the most recent frame. And I can suffer with some issues with that data stream coming in 'cause I can be tolerant with multiple sensors and processing algorithms.

Daniel Gandhi: As I send especially like one-shot messages around the system, I might say, "Oh we're gonna change modes," or a state machine changed somewhere and I want that to be conveyed to everything, I need to have some guarantee that that message is gonna get there, and that everything is gonna be able to react to it, and it can't get lost in transit.

Lacey Trebaol: This one cannot be dropped.

Daniel Gandhi: Right. And then you have some low bandwidth signals, which are on the control side, that there isn't a lot of information that is being passed in them, but they're mission-critical because they're now coupling with how the vehicle's actually moving. And so you have low bandwidth, needs to be extremely low-latency, and you can't really suffer any loss there.

Daniel Gandhi: So given that you have to manage all of these different types of information moving through the system, and you need to balance your resources appropriately, where if for whatever reason I had to sacrifice a frame of sensor data to make sure that my control signals are going out, and my one-shots are making it where they need to be, that's how it should behave. And you need to have the flexibility to push that through in the system.

Lacey Trebaol: And quality of service settings let you do that.

Daniel Gandhi: Exactly.

Lacey Trebaol: You get to actually go through and fine-tune the behaviors, so that you get the behaviors you need.

Daniel Gandhi: Yes.

Lacey Trebaol: That is a unique feature.

Daniel Gandhi: Yes.

Lacey Trebaol: Alright, have you used any of our other software services? So I know that you guys you've tried Micro before, what you guys needed was available in Pro, and now you're using Pro. Have you used like Record Replay, Routing Service?

Daniel Gandhi: We've played with it, Record and Replay, we haven't really used it extensively. I think Routing Service, we haven't done anything with it, but we're talking about connected vehicles, I think that would probably come more into play at that point, so we're not there yet.

[22:00] How Connext DDS supports AI applications

Lacey Trebaol: Okay. How does Connext support your AI applications?

Daniel Gandhi: Well, so I would say first of all, when I was talking about AI, it's a super-nebulous term, and so it's just so unbounded as to what it means. In the movement that you're seeing, is just one where you're applying machine learning and artificial intelligence to more and more complex problems. And as you apply things to more complex problems, that means you have complex hardware infrastructure, and more processes have to talk to each other. And so rather than it being like a AI or machine learning dovetails with DDS, it's more that they both fit into the same set of tools that you'd use to solve complex problems.

Daniel Gandhi: Just for context, if you were running like a neural network that were just going to like highlight objects on this table, where would DDS fit into that, right? Like it's not like you would need DDS to do that, so it's not really in that they specifically tie together, it's just that way the world is going.

[23:10] Advantages of combining machine learning and AI using deep learning

Lacey Trebaol: Alright, so I understand you're combining traditional machine learning algorithms and artificial intelligence using deep learning, can you share some of the advantages of this approach?

Daniel Gandhi: Yeah, so I would say that we are combining traditional approaches in terms of planning and control and perception, with artificial intelligence/deep learning. Machine learning is a hot topic right now, and it shows a lot of promise. The way that it essentially works is you are providing this black box a set of inputs and outputs that you want, and it is trying to find a way to regress that, so it's creating an internal model that it thinks will represent what you want from, give it a set of input to what you want it to output.

Daniel Gandhi: So that's a pretty powerful tool when you don't actually have to create the model on the inside, in terms of if this were something really trivial, you might say I can see I have some data and I want to do a linear fit to it, but I know ahead of time what a line is, and I can see that my data kind of matches a line, and I can draw the line.

Daniel Gandhi: Here we're generalizing that to the point where it doesn't know what anything needs to look like, and so it can create what it needs to model the appropriate-

Lacey Trebaol: Resulting action.

Daniel Gandhi: Exactly, so that you get the results that align with what you're looking for.

Lacey Trebaol: But how it gets there is up to it to figure out.

Daniel Gandhi: Exactly, and also your ability to inspect, to understand what it really does, is also limited, at least based on what people are able to do today. There are academics that are trying to peel back those layers and understand more of how the machine learning ends up working, but unless there's some major shift in that department, what you lack is any way to predict what the behavior's gonna be when it encounters something that it's never seen before. And they can be fairly mundane things, because it's all tied to how the model was created.

Daniel Gandhi: So you could just have some odd confluence of events, and have a neural network behave completely erratically.

Lacey Trebaol: But it's totally logical if you look at what it was operating on before, and what it's model is, right.

Daniel Gandhi: Right, and so the ability to cleanly predict what will happen means it just becomes like this game of exposure. We talked earlier about exposure, and talking about if you try really, really hard over many years, you can get up to a few million miles of exposure, and then you still have to put that through the system. You still have to have all the computation to process millions of miles of exposure.

Daniel Gandhi: And on the flip side, you're gonna put this out into the world, and it's gonna get a billion miles right off the bat. So when you talk about it from that perspective, you have to be somewhat guarded about what the machine learning algorithms are going to do. If I have some model in some aspect of a system that I could represent through physical equations, if I were saying I want to put together something that represents basically the dynamics of the vehicle, I could go and run a bunch of simulations that have a machine learning algorithm regress that, or I could try to do it analytically. And when I do it analytically, I have a better sense of the understanding of its boundaries, which means I span a space that is well-defined. And I know that if I leave this space, it won't be well-defined, and I can guard against that.

Daniel Gandhi: I don't understand the boundaries of the machine learning, aside from the fact that where did I train it, so its predictability, it's hard to create a bound on-

Lacey Trebaol: Because it's a function of its exposure?

Daniel Gandhi: Right.

Lacey Trebaol: And that's hard to put that point on.

Daniel Gandhi: Yes. So I mean, if we go back to our earlier example where we're talking about a line, if I have a set of data over a limited range and I put a line to it, everything might match really well. But if I fit a higher polynomial to it, I might still have something that looks just like a line in that range, and it might go completely erratic outside of that range. And when that isn't just a smooth space, when you have all these little holes, and you have all these corners that you have to push into, then how tractable is that? That becomes an issue where when you say I'm going to scale the way that we're talking about scaling, that you want to guard against that.

Daniel Gandhi: And so, having a blend of traditional and machine learning algorithms allows you to have safeguards. It allows you to encapsulate exactly what the responsibilities of each subsystem are, so you could say the neural network is responsible for just this, and therefore I have some sense of what ways it could be wrong and make sure that there's something else in the system that can ensure that it can cover.

Lacey Trebaol: That's a common thread in a lot of things you've worked on, needing to be very knowledgeable about two types of things in order to kind of create a hybrid solution, I guess.

Daniel Gandhi: Right.

Lacey Trebaol: Do you want to predict the future?

Daniel Gandhi: Sure.

[28:07] What does the future look like for autonomous vehicles?

Lacey Trebaol: Oh you do? I love this, this is great. So what do you think the future is going to look like for autonomous vehicles and the things that you're working on right now? You said it's half a percent of cars operating like this would be high.

Daniel Gandhi: I don't know that it would be high. I mean, I was using it as an example of just something of even in like your very conservative case of scaling like what your exposure would look like. Exactly how fast things will roll out, I think, is TBD. I would say that I don't think it will be as clear-cut as people are expecting. I think that we already have autonomy that's encroaching. If you look at something like a smartphone, I mean they were kind of like dabbling around in the Palm trio and the Blackberry days-

Lacey Trebaol: I had one of those.

Daniel Gandhi: -And then you have an iPhone come out, and that starts to become very mainstream, and I mean iPhone came out ten years ago, and you'd be loathe to find somebody who doesn't have a smartphone today. But it wasn't like there was some time when we were like ... Computing is now pushed into everybody's hands, like that wasn't like a clear line, but it happened in a manner where it's all there and we don't usually stop to reflect on it.

Lacey Trebaol: And now we don't even question it.

Daniel Gandhi: Exactly, and I think the same thing is what you're gonna see in this space. As you're looking at autonomous vehicles from the retail channel, you're just gonna start seeing more of the driving burden removed from the driver. You're still gonna-

Lacey Trebaol: And people being comfortable with that, which I think is an entirely different obstacle. Take your hands off the wheel ...

Daniel Gandhi: It's a funny thing. I remember when hybrids first came out, and people were very unsettled with the fact that the car turned off and was silent in the car, and that they had to trust that when they hit the gas that the car was gonna turn back on and you were gonna go. But I think they will become comfortable as they start seeing the convenience of it. We have had active safety out there to the point where it's becoming mainstream in commodity, with systems of more and more capability increasing over time.

Lacey Trebaol: And it hasn't increased the price of the car, by the way, on some of those measures I've noticed. So like it got into my car without adding any additional cost, and it wasn't a feature that I was seeking out, it was just oh it also comes with this. So the fact that you can bolt these things onto a car essentially, and have these additional features that people are going to get comfortable with, without increasing the price, I think will help with adoption.

Daniel Gandhi: Yes, and I would say that there are levels of capabilities, so you have-

Lacey Trebaol: Levels of autonomy.

Daniel Gandhi: Right, and exactly what that means. So I think you're starting to see mainstream vehicles are starting to have adaptive cruise control, for example. Adaptive cruise control, I think, has been around in like the luxury vehicles for a very long time, maybe like 15 years or something. As it's starting to push down into the mainstream and people get more exposure to it, they'll become more comfortable with it.

Daniel Gandhi: If you watched what the high-end is doing today in comparison, you started to see five to seven years ago, that adaptive cruise control very quietly became a full-speed range adaptive cruise, where it can take you down to a stop. It used to have some minimum speed that could operate it, and that meant now you could actually drive into a traffic jam, and the car could stop and you could be in bumper to bumper, and it would just say do we wanna go, do we wanna go, and it'll just keep going for you, and it will stop where it needs to.

Daniel Gandhi: So if you think of just what the driving burden during a traffic jam is, it's like excruciating, and humans are far more capable than any of these systems at handling these tasks, but they're also extremely mundane for us. And so all the accidents happen because we let them, and these systems are basically going to mature to the point where like, "Oh I don't wanna do this," so the car is going to do it. And there's gonna be some value add there, and I think we're gonna have this erosion, and especially as you look at like easier problems, the highway setting, you're starting to see Tesla has their autopilot system, I think Mercedes has systems out. GM just launched Super Cruise, which is their first sort of hands-free system for driving on a highway.

Lacey Trebaol: So what does hands-free on a highway mean?

Daniel Gandhi: So at least the tagline that GM has been pushing is that all of their competitors in terms of highway autonomy require that your hands are on the wheel, or that you constantly touch the wheel, so that they know that you're paying attention. What GM has done is they've put in a driver awareness system that actually tracks your eye movements. So they look to see that you're looking out the window, and if you look away for a period of time, they start to escalate to say you have to get back involved in some form, and then they'll eventually like slow down the car and whatnot.

Daniel Gandhi: What that means is that you can actually take your hands off the wheel, and the fact that you have to hold the wheel doesn't become an element anymore, and that means that as you're driving, if you like glance at different things, like as long as you don't exceed their time windows, it's starting to relieve the burden.

Daniel Gandhi: As you look at these systems and they mature, you're going to say ... here they're talking about staying in a lane and driving on a highway, what happens when I drive on the highway for a long distance. Well, maybe it will change lanes for me at some point. But I think some people have tried on some small scope, but you say okay it's gonna do an automatic lane change, now I put in my nav system and I say I wanna go from A to B, and the highway portion of that, including lane changes, as you have divides and whatnot, is covered for you. And you say okay, well now I wanna push into ramps, and then I want to push into surface streets, and it's just gonna slowly cover more aspects of-

Lacey Trebaol: And all of those future ones will actually probably be smoother because of what was learned and achieved leading up to it.

Daniel Gandhi: Exactly. You will also get a sense for more of the corner cases, right. These systems will be in the field, and you'll get some near misses that will become more critical if you have a more capable system. But you have data on that now that you can start incorporating prior to it rolling out as a major production system. So that line is, I think, going to sneak up on people. It's just gonna become a normal part of your life in a way that you don't quite think about. It'll be like some novelty to start with, but then it's just gonna be what you use every day.

Daniel Gandhi: So I think from the retail side, that that's the channel that you're see. I think on the fleet side, most likely they're gonna have to start small in terms of whether it can deploy, so they're gonna be in specific guard areas, and so maybe they'll be automated shuttle routes someplace, right, they may have their own lane to isolate something to begin with.

Lacey Trebaol: Or they have like the campus fleets?

Daniel Gandhi: Right, or like closed campus, corporate campus, just moving people around in some automated way. We see that with sort of hard infrastructure, where you see like automated trams and automated rail, so it's automating the stop and go, and then the rail keeps it steering. You're adding a steering element into it, and you can start to take away some of the environmental safeguards that you were placing around these things.

Daniel Gandhi: But at some point you get into a tram at some airport or something that's automated, people don't even like bat an eye now. And yet, somebody had to make something that could automatically without having a driver move people around-

Lacey Trebaol: And not be a danger to everyone, getting on and off it.

Daniel Gandhi: Right. You also don't hear about them failing in some dramatic way, and so I think the same thing will happen with the fleet autonomy as well. So as the technology matures, I think you're just gonna see that it kind of slowly crawls into different aspects of your life, and I don't see it being a sort of dramatic switch. And I think a lot of that is also stemming from the fact that everything is tied to some physical infrastructure. If we were gonna have a dramatic switch in the way everything operated, then we would have to change the way all the roads are structured, and what all the signals, and just the cost of that means that I don't see it happening very easily, and so that having a sort of organic, almost grassroots-buildup of the technology, is more likely to be where it comes from.

Lacey Trebaol: So I think they're normalizing aspects of it in that way, 'cause like we're hearing about it and it's available to the general consumer, but we have not all experienced it, so I feel like that lowers the boundary there a tiny bit at that point.

Lacey Trebaol: Okay here's one, they wanna know if you guys here are involved in any industry organizations or consortia? So that could mean anything from like participation with development of like the Ross Standard or something along that line, or anything public that you guys actively participate in or contribute towards somehow.

Daniel Gandhi: So I would say basically that our primary contribution in that realm has been more in, given the self-driving space is evolving, that there's some local groups that are creating a consortia between industry and academia, so we'll go to working group meetings there. Department of Transportation of Massachusetts has a working group that they have to try to address how policy regulation, how they want to register, license, how they want to insure, everything around-

Lacey Trebaol: That's an interesting point, I forgot about that. We have to insure those.

Daniel Gandhi: Yeah, so they have basically an open meeting every couple months that we'll at least attend, and maybe we'll contribute something to in terms of questions or suggestions. So I think from me personally, and what we're doing in the AV space, that's what our contribution is, and really it's around saying that, especially lawmakers and policy analysts have to find a way to cover this new space in a way that the area still stays competitive, you're not like cutting off innovation, but you don't wanna risk public safety. So there's a balancing act there, and you don't want to be in a situation where something goes wrong, and there were no regulations in place, and now there's like a steep backpedal, where you're like clamping down-

Lacey Trebaol: And they're like forget this.

Daniel Gandhi: Exactly.

Lacey Trebaol: No to all things you're doing.

Daniel Gandhi: So one of the things that makes it, I would say, favorable for industry is to have some sense of stability, so as you try to have the governmental organizations be coupled to industry organizations, it makes it so that you could say frankly that they're going be things that go wrong as time goes on, and if people are ready for it, then it's sort of a joint-

Lacey Trebaol: Ownership of the wrong?

Daniel Gandhi: Right, that you know that's something's gonna happen at some point. We're not gonna have like a knee-jerk reaction to it, we're going to talk to the public in a way that is going to say that this is isolated, these things will happen, and we're ready to-

Lacey Trebaol: But when the policy makers are working in hand with the technologists, it's not finger pointing, and one isn't lagging terribly behind to the other, so you're able to actually address those concerns in ways that don't damage what you're trying to achieve, or hinder it from progressing.

Daniel Gandhi: Right. So the problem's also broken across, like it's federal, it's state, it's local, who wants to handle what aspect of it, and there's been somewhat of a gravitation towards this wild west approach. States that do nothing, have no regulation, no oversight, that's where you want to be. But as soon as something goes wrong, as a company you might've built up infrastructure there, you might have staff there, and then suddenly that state can just flip on you, or that municipality could flip on you-

Lacey Trebaol: And then your investment is-

Daniel Gandhi: Exactly.

Lacey Trebaol: Yeah.

Daniel Gandhi: So if you feel like you need to be responsible and say we're gonna have policy and regulation that makes it so that everybody stays safe and there's a clear path forward, it also brings stability, right. You could go to a company and say you're gonna be able to operate here for 20 years while you develop this stuff because we're here to support it.

Lacey Trebaol: And also, I would look at it as while you guys, and I know this from what you said earlier, you're swamped right now with work and demos and delivering things to customers, but at the same time, you have to invest your time in these other things that aren't product.

Daniel Gandhi: Right.

Lacey Trebaol: Because it's potentially just as important to what you're trying to achieve as the actual technology, if they won't let you use it.

Daniel Gandhi: Right, and on a different side of that same coin, is supporting smaller companies. We're a smaller company.

Lacey Trebaol: How many people?

Daniel Gandhi: We're at about 35 people right now. When we go and say that we want to operate within a certain location, and you're talking about vehicles, that means that you need to have all this infrastructure in place as well. You need to have offices that are co-located or reasonably close to garage space. You need to be able to have that reasonably close to wherever you're allowed to test. If something goes wrong, like what are the procedures? If there's an accident, a local cops gonna show up, what happens?

Lacey Trebaol: What's the protocol?

Daniel Gandhi: Yeah exactly, how do you deal with all those things. And if you have a place that has very heavy-handed regulation, now you need to have like a lawyer on hand, you need to go through a legal process to get ... So basically, this sort of work as much as the small companies can be agile, just all the infrastructure requirements favor larger companies. And so if you want to have something that is a very innovative space, it kind of needs to be incubated by the government in some form. Otherwise, it's all gonna tend towards large companies, and then they're gonna set the pace of innovation.

Lacey Trebaol: Is that actually one of the reasons that you partner with larger companies? Is that part of like a model for being successful in these areas, do you think? Is there a huge advantage to that? You work with large companies that we cannot talk about.

Daniel Gandhi: Correct. Our model is to basically create strategic partnerships with larger companies. It's something that is new to them as much as we're a new organization, so it's something that we have to keep trying to navigate through, but the idea is to have a best of both worlds type of approach, where you could have a small nimble team, but backed by the resources of a larger organization. There's financial, there's facilities, if you look at what a major automotive OEM or supplier has in terms of testing facilities, they'll be like billions of dollars of investment-

Lacey Trebaol: That you could never replicate?

Daniel Gandhi: Exactly. And so, you have access to those facilities. If you look at how a lot of autonomous vehicle companies integrate into a vehicle, it's very sort of hacked on in many ways. They've tried to reverse engineer interfaces to different aspects so that they could control it, because it was never meant to be controlled in that manner.

Lacey Trebaol: Yeah, it's meant to be driven by a person.

Daniel Gandhi: Right. If you partner with a larger company that has some connectivity there, some ownership of those pieces, then now you can actually have proper interfaces, because they can go and start tweaking and changing things to make it suitable for your task.

Lacey Trebaol: So the inside of the car doesn't look like a hacked-up job, wires everywhere and all those panels removed.

Daniel Gandhi: Exactly. And if you look at what test vehicles look like coming out of suppliers and OEMs, in general, just as part of their normal development, not even for AVs, they're extremely polished compared to what you'd see other places.

Lacey Trebaol: So we talked about the fact that you're hard to find online, but you guys are having a website that by the time this podcast goes live, should be up, so do you know what that URL will be?

Daniel Gandhi: Yeah, it could or

[43:19] What happens next for NextDroid?

Lacey Trebaol: Okay, so the last question here is what happens next for NextDroid?

Daniel Gandhi: I would say that the technology for robotics and automation is broadly applicable. We are targeting some spaces right now that can make use of it today, but the where that can be applied is just going to expand over time, and so you could imagine if I have some perception algorithm for a car, well how would that apply to a UAV? How would that apply to even something just like a security system?

Daniel Gandhi: And so, we are setting ourselves up such that we can contribute to lots of industries, and as time goes on, I think you'll see that the types of vehicles that we can build or apply the technology to will expand. We'll move into more aspects of robotics like manipulation-based systems and other challenging areas along those lines.

Lacey Trebaol: Sounds like something you'd be into.

Daniel Gandhi: Oh absolutely.

Steven Onzo: Thanks for listening to Episode 16 of the Connexts Podcast, we hope you enjoyed it. Stay tuned for Part 2 of Dan's interview, where he goes into detail about safety in autonomous vehicles, as well as early projects he's worked on.

Steven Onzo: If you have any questions or suggestions for future interviews, please be sure to reach out to us on social media or Thanks and have a great day.


Get the latest updates and insights from the RTI newsletter.

Subscribe to the Newsletter