In this episode, Sabine Hauert speaks with Travis Deyle, about his IROS-nominated work on RFID tags, his blog Hizook, and the career path that brought him from academia, to founding his own start-up, and finally working for Google[x].
For his PhD at Georgia Tech with Dr. Charles C. Kemp, Deyle helped robots find household objects by tagging them with small Band-Aid-like Ultra High Frequency (UHF) Radio-Frequency Identification (RFID) labels. The tags allowed robots to precisely identify tagged objects. Once identified, the robots would follow a series of simple behaviors to navigate up to the objects and orient towards them.
Compared to vision and lasers, RFID can detect objects that are hidden while providing precise information and identification. This could allow a robot to find a bottle of medication in a cupboard, and make sure it’s the correct medication, before bringing it to a person. Furthermore, the technology can scale to large numbers of objects, and be used to map their location in the environment.
In the future, such tags augmented with better energy, sensing and computation capabilities could form the basis of the Internet of Things and provide a smart environment for robots to interact with.
Travis Deyle earned a PhD in Fall 2011 from Georgia Tech’s School of Electrical and Computer Engineering (ECE). His PhD with Dr. Charles C. Kemp at the at Healthcare Robotics Lab was entitled, “Ultra High Frequency (UHF) Radio-Frequency Identification (RFID) for Robot Perception and Mobile Manipulation.”
After his PhD, Deyle worked with Dr. Matt Reynolds as a postdoc researcher at Duke University where he focused on a software-defined radio receiver to decode (in real-time) the high-speed biotelemetry signals reflected by a custom neuro-telemetry chip. This system was designed to capture high-fidelity neural signals from a dragonfly in flight — aka, a “cyborg dragonfly”.
He then co-founded the successful company Lollipuff.com: an online auction site dedicated exclusively to women’s designer clothes and accessories.
Deyle currently works at Google[x] where he was part of the team that made the “smart contact lense” to measure tear glucose levels which was recently licensed to Novartis.
He also founded the well know blog Hizook.com, a robotics website for academic and professional roboticists.
Sabine: Hi Travis, welcome to robots.
Travis: Thanks Sabine.
Sabine: Do you want to start by introducing yourself to our listeners?
Travis: My name is Travis Deyle, and I did my PhD at Georgia Tech in the Healthcare Robotics Lab where I worked with a lot of large mobile manipulating robots. After Georgia Tech I headed up to Duke University, did a Post-Doc, and now I’m working at Google[x].
Sabine: That’s a dream trajectory for a lot of young roboticists! You were nominated for Best Paper Award here at IROS … do you want to tell us a little bit more about your research?
Travis: Absolutely. The research we’re presenting [at IROS] is about affixing very small labels to various household objects … things you would typically find in your home like a TV remote, a hairbrush, toys and medication … and by attaching these Band-Aid-like tags, we’re able to get precise identity information about various objects.
The problem is that the tags alone do not provide location information, so we’ve developed a series of very simple robot behaviors that allow a robot to navigate up to the tags and orient towards them. It’s not actually localizing the tag per se, but getting very close to it so that you can bring other sensors to bear. Compared to vision and lasers, this is really neat because: 1) you don’t need line of sight; 2) it provides precise identity; and 3) it can work in cases where there might be a large risk of you didn’t get the identification right, for example with medication bottles.
Sabine: So you actually have one of these tags in your hand. Can you describe it for us?
Travis: The tag is about 10 cm long by 1 cm wide, and it literally looks like a Band-Aid in that it’s super thin. It’s basically a piece of paper with a very small electric integrated circuit chip sitting in the middle. It’s mostly antennae, and it has a very tiny integrated circuit that provides the identification. These tags have no battery – they’re long range RFID – so they operate at about 900 MHz and can be read from about six meters away when you have a perfect line of sight. More practically, when you start tagging objects in a home, you get maybe a three meter range. So it’s a bit different from the classic RFID you might have in your wallet or for access control.
Sabine: How does the robot go and find these tags?
Travis: The behaviors are simple. They rely on having a directional antenna, which you can think of as being like a metal detector of sorts. This antenna has a beam, and the robot moves around and points these antennae in different directions, and all it’s doing is looking for where it gets the strongest signal.
We came up with three behaviors that we use. First, the robot can traverse around the room and look for all the tags in the environment. Second (once you know what object you want), the robot goes back to where it saw that tag and it pans the antenna around and orients itself in the direction of the strongest signal. The third behavior is to use two different antennae and to servo based on the different signal, and then the robot will continue moving until it gets obstructed by an obstacle.
What we found in our paper is that these three simple behaviors can be combined to allow the robot to find the tag and then navigate up to it. In fact we have unpublished results that actually compare these to probabilistic techniques, and for the objective measures that we use, it’s either on par or better than using the probabilistic techniques alone. Of course you can imagine combining them to get even better results.
Sabine: There’s been quite a few techniques to go up gradients with robots, and I wonder what makes it difficult with RFID?
Travis: That’s actually a very good point. These are very classic robot behavior-based techniques. RFID is unique in the sense that there are a whole lot of things that can dramatically change the signals you get from tags. You have classic things like multipath where signals bounce of walls, you have other things like shadowing where maybe I’m blocking the signal because there’s something in the way. But RFID actually has this very troublesome aspect to it, in that if you start tagging objects, now you have objects in lots of different poses and around lots of different other objects, and the simple act of taking a tagged object and moving it to a new location can dramatically change the RF signals that you receive. That’s something that’s very hard to capture in these Big-Data-driven models ahead of time. The models end up having to be so generic that you lose a lot of the discrimination capabilities of the model.
Sabine: Do the robots map as they go, and basically remember where these RFIDs are? They find them and then that information is stored somewhere?
Travis: That’s exactly right. The robot has a map of, say your home, and as it’s going round your house doing all these other (hopefully useful) robot activities, it’s able to keep track of where it last obtained readings from the tags. That way when you query and say, “Hey robot, go find the tag,” it knows to go back to roughly the location that it got the last positive readings.
Sabine: Why go for that rather than vision or lasers?
Travis: I actually think it’s complimentary. I don’t think RFID is going to replace vision and lasers. In fact, if you can use vision to do it, that’s great. These techniques are really well suited for circumstances where you lost an object, or if something is hidden out of sight. Then the robot get positive signals from exactly the item you want ahead of time. Medication is [another case where this could be useful]. It would be very powerful if robots could deliver the right medication to the right person at the right time, but if you’ve ever looked inside of a medication cabinet, you’d see that a lot of the bottles look very similar. [RFID] gives the robot a way to precisely identify the correct objects even among visually similar ones. When you’re dealing with something like medication adherence, the consequences of having a false positive are really, really dire.
Sabine: It sounds like this is something that’s almost ready for market … is that the case?
Travis: I think there’s a lot of potential to use RFID to bootstrap into marketplaces. I don’t know of any commercial systems right now that are using it beyond classic inventory tracking in warehouses. What we hope is that, while these methods may not enable a robot to be immediately commercialized, they will hasten that process so that it can happen much sooner than it otherwise would.
Sabine: There are so many objects that you’d want robots to manipulate in your everyday home. Does this scale to lots of RFIDs?
Travis: Yeah, so these tags they’re a commercial standard. [The robot] can query for hundred or thousands of tags in the same environment, and it can do this one of two modes. In one mode, it says “Every tag out there: please respond,” and it has some probabilistic way in which it basically uses that channel at the same time. And that’s built into the standard, so we didn’t have to invent anything for that to be the case. In the other mode, the robot or the reader would say, “I’m looking for this specific tag. Can only this tag chime up and say that’s it’s there?”
As we start tagging objects around the home, it’s almost like giving explicit permission to a robot to actually go and interact with those items. So if you’ve got things in your house that you don’t want the robot to touch, this might be a good way to tell it which ones are okay.
Sabine: I want to also ask you about science communication, because your blog Hizook is very well-known in the robotics world. What got you there, why did you start it and where is it going?
Travis: Hizook started because I had all these interesting robotics ideas that I wanted to get out there, and it was a way to get them out of my head. Once those ideas are on paper, then you have something you can point to. I wish I could say it started because of some big altruistic idea, but in reality it was just my own way of jotting down thoughts to get them on paper.
Sabine: Would you recommend it to other young researchers?
Travis: Absolutely. I think science communication is really important. It becomes necessary to be able to communicate your results to a lay audience in a way that they can understand both the amazing capabilities that you’ve provided to robots, as well as the limitations. I think everyone should start a website, or at least contribute articles, in such a way that they can communicate their results.
Sabine: Speaking of limitations, what are some of the limitations of your RFID tags?
Travis: These RFID tags, the very low cost ones, are about 10 cents apiece and they work really great on things like plastics and cardboard. However, if you add metal it interferes with the RF propagation of the signals, and so they have to make special “on-metal” tags. These tags can be more expensive – a dollar or so per tag. I have one on my keys actually.
Sabine: I would need one on my keys … I would need one on everything that I own!
Travis: [The other limitation is that] you don’t get precise pose information. In practice you’d probably want to combine this with vision, laser depth cameras or even short range RFID, which is some of the other work we did. You can imagine using long range RFID from afar, and then when you get up close using short range. There are certainly some limitations.
Sabine: What are the next steps?
Travis: I think you’re going to see that UHF tags are going to become more and more common. In many ways they are a form of Internet of Things, in that you have this embodied intelligence sitting on a device. The next set of tags is not going to be simply for identification – they’ll actually have sensing on board. We’re starting to see those tags start to come out commercially now. I think there are some very interesting possibilities for tags: they’re battery free, they harvest all of their energy, and they have general-purpose computation, general-purpose sensing on board, and you can start embedding them in every type of device out there for basically no cost.
Sabine: You have a trajectory that a lot of young researchers would like to have, having begun in the academic world, then you founded a successful startup, and then went to Google. So you’ve seen a lot of things that other people are hesitating about. They’re wondering: should I go to industry, should I go to startup, should I do academia? Do you have any tips? Are you happy with what you did, would you change it? Why?
Travis: I’m pretty happy with it. I would say I’ve been opportunistic throughout the whole process. Sadly I’m not dogmatic about just robotics. I think a lot of fundamental robotic technologies have potential to be startups on their own. Some examples might even be depth cameras, and structure-for-motion mapping. We’re starting to see a lot of products outside the robotics realm that are actually using these and being very successful, and providing a lot of value. I think robotics is great because it pulls together everything and it also has this potential to explode into separate little entities. I’m sure that someday I’ll come back to robotics … because with sensing, actuation, perception … you get a little bit of everything.
Sabine: So the tip would be to be opportunistic? Go for it?
Travis: I think so, yeah. And be aware of the opportunities that are out there. Following your passion, whatever that might be, is a pretty significant part of that.
Sabine: Since you have a big view of the robotics field because of your blog, where do you see this going? Do you have any insight on the future of robotics?
Travis: It’s hard to say. I think if there was something that was immediately obvious in robotics, I’d probably be doing that right now. The real key is to focus on applications of robotics that can provide value, and that doesn’t necessarily mean making sexy robots because they’re awesome. A lot of it has to do with being in the trenches, finding problems that lend themselves to automation, and then providing value by building automated solutions. It’s kind of a chicken-and-egg problem, and I think that, coming out of academia, I focused for a long time on building cool robots. Now I’m taking the opposite approach, which is going and looking at applications, and hopefully there will be something that ties back to robotics on the flip side.
Sabine: All right, thanks Travis for being here with us on robots!
Travis: No problem!All audio interviews are transcribed and edited for clarity with great care, however, we cannot assume responsibility for their accuracy.