Tag Archives: huge

#437783 Ex-Googler’s Startup Comes Out of ...

Over the last 10 years, the PR2 has helped roboticists make an enormous amount of progress in mobile manipulation over a relatively short time. I mean, it’s been a decade already, but still—robots are hard, and giving a bunch of smart people access to a capable platform where they didn’t have to worry about hardware and could instead focus on doing interesting and useful things helped to establish a precedent for robotics research going forward.

Unfortunately, not everyone can afford an enormous US $400,000 robot, and even if they could, PR2s are getting very close to the end of their lives. There are other mobile manipulators out there taking the place of the PR2, but so far, size and cost have largely restricted them to research labs. Lots of good research is being done, but it’s getting to the point where folks want to take the next step: making mobile manipulators real-world useful.

Today, a company called Hello Robot is announcing a new mobile manipulator called the Stretch RE1. With offices in the San Francisco Bay Area and in Atlanta, Ga., Hello Robot is led by Aaron Edsinger and Charlie Kemp, and by combining decades of experience in industry and academia they’ve managed to come up with a robot that’s small, lightweight, capable, and affordable, all at the same time. For now, it’s a research platform, but eventually, its creators hope that it will be able to come into our homes and take care of us when we need it to.

A fresh look at mobile manipulators
To understand the concept behind Stretch, it’s worth taking a brief look back at what Edsinger and Kemp have been up to for the past 10 years. Edsinger co-founded Meka Robotics in 2007, which built expensive, high performance humanoid arms, torsos, and heads for the research market. Meka was notable for being the first robotics company (as far as we know) to sell robot arms that used series elastic actuators, and the company worked extensively with Georgia Tech researchers. In 2011, Edsinger was one of the co-founders of Redwood Robotics (along with folks from SRI and Willow Garage), which was going to develop some kind of secret and amazing new robot arm before Google swallowed it in late 2013. At the same time, Google also acquired Meka and a bunch of other robotics companies, and Edsinger ended up at Google as one of the directors of its robotics program, until he left to co-found Hello Robot in 2017.

Meanwhile, since 2007 Kemp has been a robotics professor at Georgia Tech, where he runs the Healthcare Robotics Lab. Kemp’s lab was one of the 11 PR2 beta sites, giving him early experience with a ginormous mobile manipulator. Much of the research that Kemp has spent the last decade on involves robots providing assistance to untrained users, often through direct physical contact, and frequently either in their own homes or in a home environment. We should mention that the Georgia Tech PR2 is still going, most recently doing some clever material classification work in a paper for IROS later this year.

Photo: Hello Robot

Hello Robot co-founder and CEO Aaron Edsinger says that, although Stretch is currently a research platform, he hopes to see the robot deployed in home environments, adding that the “impact we want to have is through robots that are helpful to people in society.”

So with all that in mind, where’d Hello Robot come from? As it turns out, both Edsinger and Kemp were in Rodney Brooks’ group at MIT, so it’s perhaps not surprising that they share some of the same philosophies about what robots should be and what they should be used for. After collaborating on a variety of projects over the years, in 2017 Edsinger was thinking about his next step after Google when Kemp stopped by to show off some video of a new robot prototype that he’d been working on—the prototype for Stretch. “As soon as I saw it, I knew that was exactly the kind of thing I wanted to be working on,” Edsinger told us. “I’d become frustrated with the complexity of the robots being built to do manipulation in home environments and around people, and it solved a lot of problems in an elegant way.”

For Kemp, Stretch is an attempt to get everything he’s been teaching his robots out of his lab at Georgia Tech and into the world where it can actually be helpful to people. “Right from the beginning, we were trying to take our robots out to real homes and interact with real people,” says Kemp. Georgia Tech’s PR2, for example, worked extensively with Henry and Jane Evans, helping Henry (a quadriplegic) regain some of the bodily autonomy he had lost. With the assistance of the PR2, Henry was able to keep himself comfortable for hours without needing a human caregiver to be constantly with him. “I felt like I was making a commitment in some ways to some of the people I was working with,” Kemp told us. “But 10 years later, I was like, where are these things? I found that incredibly frustrating. Stretch is an effort to try to push things forward.”

A robot you can put in the backseat of a car
One way to put Stretch in context is to think of it almost as a reaction to the kitchen sink philosophy of the PR2. Where the PR2 was designed to be all the robot anyone could ever need (plus plenty of robot that nobody really needed) embodied in a piece of hardware that weighs 225 kilograms and cost nearly half a million dollars, Stretch is completely focused on being just the robot that is actually necessary in a form factor that’s both much smaller and affordable. The entire robot weighs a mere 23 kg in a footprint that’s just a 34 cm square. As you can see from the video, it’s small enough (and safe enough) that it can be moved by a child. The cost? At $17,950 apiece—or a bit less if you buy a bunch at once—Stretch costs a fraction of what other mobile manipulators sell for.

It might not seem like size or weight should be that big of an issue, but it very much is, explains Maya Cakmak, a robotics professor at the University of Washington, in Seattle. Cakmak worked with PR2 and Henry Evans when she was at Willow Garage, and currently has access to both a PR2 and a Fetch research robot. “When I think about my long term research vision, I want to deploy service robots in real homes,” Cakmak told us. Unfortunately, it’s the robots themselves that have been preventing her from doing this—both the Fetch and the PR2 are large enough that moving them anywhere requires a truck and a lift, which also limits the home that they can be used in. “For me, I felt immediately that Stretch is very different, and it makes a lot of sense,” she says. “It’s safe and lightweight, you can probably put it in the backseat of a car.” For Cakmak, Stretch’s size is the difference between being able to easily take a robot to the places she wants to do research in, and not. And cost is a factor as well, since a cheaper robot means more access for her students. “I got my refurbished PR2 for $180,000,” Cakmak says. “For that, with Stretch I could have 10!”

“I felt immediately that Stretch is very different. It’s safe and lightweight, you can probably put it in the backseat of a car. I got my refurbished PR2 for $180,000. For that, with Stretch I could have 10!”
—Maya Cakmak, University of Washington

Of course, a portable robot doesn’t do you any good if the robot itself isn’t sophisticated enough to do what you need it to do. Stretch is certainly a compromise in functionality in the interest of small size and low cost, but it’s a compromise that’s been carefully thought out, based on the experience that Edsinger has building robots and the experience that Kemp has operating robots in homes. For example, most mobile manipulators are essentially multi-degrees-of-freedom arms on mobile bases. Stretch instead leverages its wheeled base to move its arm in the horizontal plane, which (most of the time) works just as well as an extra DoF or two on the arm while saving substantially on weight and cost. Similarly, Stretch relies almost entirely on one sensor, an Intel RealSense D435i on a pan-tilt head that gives it a huge range of motion. The RealSense serves as a navigation camera, manipulation camera, a 3D mapping system, and more. It’s not going to be quite as good for a task that might involve fine manipulation, but most of the time it’s totally workable and you’re saving on cost and complexity.

Stretch has been relentlessly optimized to be the absolutely minimum robot to do mobile manipulation in a home or workplace environment. In practice, this meant figuring out exactly what it was absolutely necessary for Stretch to be able to do. With an emphasis on manipulation, that meant defining the workspace of the robot, or what areas it’s able to usefully reach. “That was one thing we really had to push hard on,” says Edsinger. “Reachability.” He explains that reachability and a small mobile base tend not to go together, because robot arms (which tend to weigh a lot) can cause a small base to tip, especially if they’re moving while holding a payload. At the same time, Stretch needed to be able to access both countertops and the floor, while being able to reach out far enough to hand people things without having to be right next to them. To come up with something that could meet all those requirements, Edsinger and Kemp set out to reinvent the robot arm.

Stretch’s key innovation: a stretchable arm
The design they came up with is rather ingenious in its simplicity and how well it works. Edsinger explains that the arm consists of five telescoping links: one fixed and four moving. They are constructed of custom carbon fiber, and are driven by a single motor, which is attached to the robot’s vertical pole. The strong, lightweight structure allows the arm to extend over half a meter and hold up to 1.5 kg. Although the company has a patent pending for the design, Edsinger declined to say whether the links are driven by a belt, cables, or gears. “We don’t want to disclose too much of the secret sauce [with regard to] the drive mechanism.” He added that the arm was “one of the most significant engineering challenges on the robot in terms of getting the desired reach, compactness, precision, smoothness, force sensitivity, and low cost to all happily coexist.”

Photo: Hello Robot

Stretch’s arm consists of five telescoping links constructed of custom carbon fiber, and are driven by a single motor, which is attached to the robot’s vertical pole, minimizing weight and inertia. The arm has a reach of over half a meter and can hold up to 1.5 kg.

Another interesting features of Stretch is its interface with the world—its gripper. There are countless different gripper designs out there, each and every one of which is the best at gripping some particular subset of things. But making a generalized gripper for all of the stuff that you’d find in a home is exceptionally difficult. Ideally, you’d want some sort of massive experimental test program where thousands and thousands of people test out different gripper designs in their homes for long periods of time and then tell you which ones work best. Obviously, that’s impractical for a robotics startup, but Kemp realized that someone else was already running the study for him: Amazon.

“I had this idea that there are these assistive grabbers that people with disabilities use to grasp objects in the real world,” he told us. Kemp went on Amazon’s website and looked at the top 10 grabbers and the reviews from thousands of users. He then bought a bunch of different ones and started testing them. “This one [Stretch’s gripper], I almost didn’t order it, it was such a weird looking thing,” he says. “But it had great reviews on Amazon, and oh my gosh, it just blew away the other grabbers. And I was like, that’s it. It just works.”

Stretch’s teleoperated and autonomous capabilities
As with any robot intended to be useful outside of a structured environment, hardware is only part of the story, and arguably not even the most important part. In order for Stretch to be able to operate out from under the supervision of a skilled roboticist, it has to be either easy to control, or autonomous. Ideally, it’s both, and that’s what Hello Robot is working towards, although things didn’t start out that way, Kemp explains. “From a minimalist standpoint, we began with the notion that this would be a teleoperated robot. But in the end, you just don’t get the real power of the robot that way, because you’re tied to a person doing stuff. As much as we fought it, autonomy really is a big part of the future for this kind of system.”

Here’s a look at some of Stretch’s teleoperated capabilities. We’re told that Stretch is very easy to get going right out of the box, although this teleoperation video from Hello Robot looks like it’s got a skilled and experienced user in the loop:

For such a low-cost platform, the autonomy (even at this early stage) is particularly impressive:

Since it’s not entirely clear from the video exactly what’s autonomous, here’s a brief summary of a couple of the more complex behaviors that Kemp sent us:

Object grasping: Stretch uses its 3D camera to find the nearest flat surface using a virtual overhead view. It then segments significant blobs on top of the surface. It selects the largest blob in this virtual overhead view and fits an ellipse to it. It then generates a grasp plan that makes use of the center of the ellipse and the major and minor axes. Once it has a plan, Stretch orients its gripper, moves to the pre-grasp pose, moves to the grasp pose, closes its gripper based on the estimated object width, lifts up, and retracts.
Mapping, navigating, and reaching to a 3D point: These demonstrations all use FUNMAP (Fast Unified Navigation, Manipulation and Planning). It’s all novel custom Python code. Even a single head scan performed by panning the 3D camera around can result in a very nice 3D representation of Stretch’s surroundings that includes the nearby floor. This is surprisingly unusual for robots, which often have their cameras too low to see many interesting things in a human environment. While mapping, Stretch selects where to scan next in a non-trivial way that considers factors such as the quality of previous observations, expected new observations, and navigation distance. The plan that Stretch uses to reach the target 3D point has been optimized for navigation and manipulation. For example, it finds a final robot pose that provides a large manipulation workspace for Stretch, which must consider nearby obstacles, including obstacles on the ground.
Object handover: This is a simple demonstration of object handovers. Stretch performs Cartesian motions to move its gripper to a body-relative position using a good motion heuristic, which is to extend the arm as the last step. These simple motions work well due to the design of Stretch. It still surprises me how well it moves the object to comfortable places near my body, and how unobtrusive it is. The goal point is specified relative to a 3D frame attached to the person’s mouth estimated using deep learning models (shown in the RViz visualization video). Specifically, Stretch targets handoff at a 3D point that is 20 cm below the estimated position of the mouth and 25 cm away along the direction of reaching.

Much of these autonomous capabilities come directly from Kemp’s lab, and the demo code is available for anyone to use. (Hello Robot says all of Stretch’s software is open source.)

Photo: Hello Robot

Hello Robot co-founder and CEO Aaron Edsinger says Stretch is designed to work with people in homes and workplaces and can be teleoperated to do a variety of tasks, including picking up toys, removing laundry from a dryer, and playing games with kids.

As of right now, Stretch is very much a research platform. You’re going to see it in research labs doing research things, and hopefully in homes and commercial spaces as well, but still under the supervision of professional roboticists. As you may have guessed, though, Hello Robot’s vision is a bit broader than that. “The impact we want to have is through robots that are helpful to people in society,” Edsinger says. “We think primarily in the home context, but it could be in healthcare, or in other places. But we really want to have our robots be impactful, and useful. To us, useful is exciting.” Adds Kemp: “I have a personal bias, but we’d really like this technology to benefit older adults and caregivers. Rather than creating a specialized assistive device, we want to eventually create an inexpensive consumer device for everyone that does lots of things.”

Neither Edsinger nor Kemp would say much more on this for now, and they were very explicit about why—they’re being deliberately cautious about raising expectations, having seen what’s happened to some other robotics companies over the past few years. Without VC funding (Hello Robot is currently bootstrapping itself into existence), Stretch is being sold entirely on its own merits. So far, it seems to be working. Stretch robots are already in a half dozen research labs, and we expect that with today’s announcement, we’ll start seeing them much more frequently.

This article appears in the October 2020 print issue as “A Robot That Keeps It Simple.” Continue reading

Posted in Human Robots

#437778 A Bug-Sized Camera for Bug-Sized Robots ...

As if it’s not hard enough to make very small mobile robots, once you’ve gotten the power and autonomy all figured out (good luck with that), your robot isn’t going to be all that useful unless it can carry some payload. And the payload that everybody wants robots to carry is a camera, which is of course a relatively big, heavy, power hungry payload. Great, just great.

This whole thing is frustrating because tiny, lightweight, power efficient vision systems are all around us. Literally, all around us right this second, stuffed into the heads of insects. We can’t make anything quite that brilliant (yet), but roboticists from the University of Washington, in Seattle, have gotten us a bit closer, with the smallest wireless, steerable video camera we’ve ever seen—small enough to fit on the back of a microbot, or even a live bug.

To make a camera this small, the UW researchers, led by Shyam Gollakota, a professor of computer science and engineering, had to start nearly from scratch, primarily because existing systems aren’t nearly so constrained by power availability. Even things like swallowable pill cameras require batteries that weigh more than a gram, but only power the camera for under half an hour. With a focus on small size and efficiency, they started with an off-the-shelf ultra low-power image sensor that’s 2.3 mm wide and weighs 6.7 mg. They stuck on a Bluetooth 5.0 chip (3 mm wide, 6.8 mg), and had a fun time connecting those two things together without any intermediary hardware to broadcast the camera output. A functional wireless camera also requires a lens (20 mg) and an antenna, which is just 5 mm of wire. An accelerometer is useful so that insect motion can be used to trigger the camera, minimizing the redundant frames that you’d get from a robot or an insect taking a nap.

Photo: University of Washington

The microcamera developed by the UW researchers can stream monochrome video at up to 5 frames per second to a cellphone 120 meters away.

The last bit to make up this system is a mechanically steerable “head,” weighing 35 mg and bringing the total weight of the wireless camera system to 84 mg. If the look of the little piezoelectric actuator seems familiar, you have very good eyes because it’s tiny, and also, it’s the same kind of piezoelectric actuator that the folks at UW use to power their itty bitty flying robots. It’s got a 60-degree panning range, but also requires a 96 mg boost converter to function, which is a huge investment in size and weight just to be able to point the camera a little bit. But overall, the researchers say that this pays off, because not having to turn the entire robot (or insect) when you want to look around reduces the energy consumption of the system as a whole by a factor of up to 84 (!).

Photo: University of Washington

Insects are very mobile platforms for outdoor use, but they’re also not easy to steer, so the researchers also built a little insect-scale robot that they could remotely control while watching the camera feed. As it turns out, this seems to be the smallest, power-autonomous terrestrial robot with a camera ever made.

This efficiency means that the wireless camera system can stream video frames (160×120 pixels monochrome) to a cell phone up to 120 meters away for up to 6 hours when powered by a 0.5-g, 10-mAh battery. A live, first-bug view can be streamed at up to 5 frames per second. The system was successfully tested on a pair of darkling beetles that were allowed to roam freely outdoors, and the researchers noted that they could also mount it on spiders or moths, or anything else that could handle the payload. (The researchers removed the electronics from the insects after the experiments and observed no noticeable adverse effects on their behavior.)

The researchers are already thinking about what it might take to put a wireless camera system on something that flies, and it’s not going to be easy—a bumblebee can only carry between 100 and 200 mg. The power system is the primary limitation here, but it might be possible to use a solar cell to cut down on battery requirements. And the camera itself could be scaled down as well, by using a completely custom sensor and a different type of lens. The other thing to consider is that with a long-range wireless link and a vision system, it’s possible to add sophisticated vision-based autonomy to tiny robots by doing the computation remotely. So, next time you see something scuttling across the ground, give it another look, because it might be looking right back at you.

“Wireless steerable vision for live insects and insect-scale robots,” by Vikram Iyer, Ali Najafi, Johannes James, Sawyer Fuller, and Shyamnath Gollakota from the University of Washington, is published in Science Robotics. Continue reading

Posted in Human Robots

#437769 Q&A: Facebook’s CTO Is at War With ...

Photo: Patricia de Melo Moreira/AFP/Getty Images

Facebook chief technology officer Mike Schroepfer leads the company’s AI and integrity efforts.

Facebook’s challenge is huge. Billions of pieces of content—short and long posts, images, and combinations of the two—are uploaded to the site daily from around the world. And any tiny piece of that—any phrase, image, or video—could contain so-called bad content.

In its early days, Facebook relied on simple computer filters to identify potentially problematic posts by their words, such as those containing profanity. These automatically filtered posts, as well as posts flagged by users as offensive, went to humans for adjudication.

In 2015, Facebook started using artificial intelligence to cull images that contained nudity, illegal goods, and other prohibited content; those images identified as possibly problematic were sent to humans for further review.

By 2016, more offensive photos were reported by Facebook’s AI systems than by Facebook users (and that is still the case).

In 2018, Facebook CEO Mark Zuckerberg made a bold proclamation: He predicted that within five or ten years, Facebook’s AI would not only look for profanity, nudity, and other obvious violations of Facebook’s policies. The tools would also be able to spot bullying, hate speech, and other misuse of the platform, and put an immediate end to them.

Today, automated systems using algorithms developed with AI scan every piece of content between the time when a user completes a post and when it is visible to others on the site—just fractions of a second. In most cases, a violation of Facebook’s standards is clear, and the AI system automatically blocks the post. In other cases, the post goes to human reviewers for a final decision, a workforce that includes 15,000 content reviewers and another 20,000 employees focused on safety and security, operating out of more than 20 facilities around the world.

In the first quarter of this year, Facebook removed or took other action (like appending a warning label) on more than 9.6 million posts involving hate speech, 8.6 million involving child nudity or exploitation, almost 8 million posts involving the sale of drugs, 2.3 million posts involving bullying and harassment, and tens of millions of posts violating other Facebook rules.

Right now, Facebook has more than 1,000 engineers working on further developing and implementing what the company calls “integrity” tools. Using these systems to screen every post that goes up on Facebook, and doing so in milliseconds, is sucking up computing resources. Facebook chief technology officer Mike Schroepfer, who is heading up Facebook’s AI and integrity efforts, spoke with IEEE Spectrum about the team’s progress on building an AI system that detects bad content.

Since that discussion, Facebook’s policies around hate speech have come under increasing scrutiny, with particular attention on divisive posts by political figures. A group of major advertisers in June announced that they would stop advertising on the platform while reviewing the situation, and civil rights groups are putting pressure on others to follow suit until Facebook makes policy changes related to hate speech and groups that promote hate, misinformation, and conspiracies.

Facebook CEO Mark Zuckerberg responded with news that Facebook will widen the category of what it considers hateful content in ads. Now the company prohibits claims that people from a specific race, ethnicity, national origin, religious affiliation, caste, sexual orientation, gender identity, or immigration status are a threat to the physical safety, health, or survival of others. The policy change also aims to better protect immigrants, migrants, refugees, and asylum seekers from ads suggesting these groups are inferior or expressing contempt. Finally, Zuckerberg announced that the company will label some problematic posts by politicians and government officials as content that violates Facebook’s policies.

However, civil rights groups say that’s not enough. And an independent audit released in July also said that Facebook needs to go much further in addressing civil rights concerns and disinformation.

Schroepfer indicated that Facebook’s AI systems are designed to quickly adapt to changes in policy. “I don’t expect considerable technical changes are needed to adjust,” he told Spectrum.

This interview has been edited and condensed for clarity.

IEEE Spectrum: What are the stakes of content moderation? Is this an existential threat to Facebook? And is it critical that you deal well with the issue of election interference this year?

Schroepfer: It’s probably existential; it’s certainly massive. We are devoting a tremendous amount of our attention to it.

The idea that anyone could meddle in an election is deeply disturbing and offensive to all of us here, just as people and citizens of democracies. We don’t want to see that happen anywhere, and certainly not on our watch. So whether it’s important to the company or not, it’s important to us as people. And I feel a similar way on the content-moderation side.

There are not a lot of easy choices here. The only way to prevent people, with certainty, from posting bad things is to not let them post anything. We can take away all voice and just say, “Sorry, the Internet’s too dangerous. No one can use it.” That will certainly get rid of all hate speech online. But I don’t want to end up in that world. And there are variants of that world that various governments are trying to implement, where they get to decide what’s true or not, and you as a person don’t. I don’t want to get there either.

My hope is that we can build a set of tools that make it practical for us to do a good enough job, so that everyone is still excited about the idea that anyone can share what they want, and so that Facebook is a safe and reasonable place for people to operate in.

Spectrum: You joined Facebook in 2008, before AI was part of the company’s toolbox. When did that change? When did you begin to think that AI tools would be useful to Facebook?

Schroepfer: Ten years ago, AI wasn’t commercially practical; the technology just didn’t work very well. In 2012, there was one of those moments that a lot of people point to as the beginning of the current revolution in deep learning and AI. A computer-vision model—a neural network—was trained using what we call supervised training, and it turned out to be better than all the existing models.

Spectrum: How is that training done, and how did computer-vision models come to Facebook?

Image: Facebook

Just Broccoli? Facebook’s image analysis algorithms can tell the difference between marijuana [left] and tempura broccoli [right] better than some humans.

Schroepfer: Say I take a bunch of photos and I have people look at them. If they see a photo of a cat, they put a text label that says cat; if it’s one of a dog, the text label says dog. If you build a big enough data set and feed that to the neural net, it learns how to tell the difference between cats and dogs.

Prior to 2012, it didn’t work very well. And then in 2012, there was this moment where it seemed like, “Oh wow, this technique might work.” And a few years later we were deploying that form of technology to help us detect problematic imagery.

Spectrum: Do your AI systems work equally well on all types of prohibited content?

Schroepfer: Nudity was technically easiest. I don’t need to understand language or culture to understand that this is either a naked human or not. Violence is a much more nuanced problem, so it was harder technically to get it right. And with hate speech, not only do you have to understand the language, it may be very contextual, even tied to recent events. A week before the Christchurch shooting [New Zealand, 2019], saying “I wish you were in the mosque” probably doesn’t mean anything. A week after, that might be a terrible thing to say.

Spectrum: How much progress have you made on hate speech?

Schroepfer: AI, in the first quarter of 2020, proactively detected 88.8 percent of the hate-speech content we removed, up from 80.2 percent in the previous quarter. In the first quarter of 2020, we took action on 9.6 million pieces of content for violating our hate-speech policies.

Image: Facebook

Off Label: Sometimes image analysis isn’t enough to determine whether a picture posted violates the company’s policies. In considering these candy-colored vials of marijuana, for example, the algorithms can look at any accompanying text and, if necessary, comments on the post.

Spectrum: It sounds like you’ve expanded beyond tools that analyze images and are also using AI tools that analyze text.

Schroepfer: AI started off as very siloed. People worked on language, people worked on computer vision, people worked on video. We’ve put these things together—in production, not just as research—into multimodal classifiers.

[Schroepfer shows a photo of a pan of Rice Krispies treats, with text referring to it as a “potent batch”] This is a case in which you have an image, and then you have the text on the post. This looks like Rice Krispies. On its own, this image is fine. You put the text together with it in a bigger model; that can then understand what’s going on. That didn’t work five years ago.

Spectrum: Today, every post that goes up on Facebook is immediately checked by automated systems. Can you explain that process?

Image: Facebook

Bigger Picture: Identifying hate speech is often a matter of context. Either the text or the photo in this post isn’t hateful standing alone, but putting them together tells a different story.

Schroepfer: You upload an image and you write some text underneath it, and the systems look at both the image and the text to try to see which, if any, policies it violates. Those decisions are based on our Community Standards. It will also look at other signals on the posts, like the comments people make.

It happens relatively instantly, though there may be times things happen after the fact. Maybe you uploaded a post that had misinformation in it, and at the time you uploaded it, we didn’t know it was misinformation. The next day we fact-check something and scan again; we may find your post and take it down. As we learn new things, we’re going to go back through and look for violations of what we now know to be a problem. Or, as people comment on your post, we might update our understanding of it. If people are saying, “That’s terrible,” or “That’s mean,” or “That looks fake,” those comments may be an interesting signal.

Spectrum: How is Facebook applying its AI tools to the problem of election interference?

Schroepfer: I would split election interference into two categories. There are times when you’re going after the content, and there are times you’re going after the behavior or the authenticity of the person.

On content, if you’re sharing misinformation, saying, “It’s super Wednesday, not super Tuesday, come vote on Wednesday,” that’s a problem whether you’re an American sitting in California or a foreign actor.

Other times, people create a series of Facebook pages pretending they’re Americans, but they’re really a foreign entity. That is a problem on its own, even if all the content they’re sharing completely meets our Community Standards. The problem there is that you have a foreign government running an information operation.

There, you need different tools. What you’re trying to do is put pieces together, to say, “Wait a second. All of these pages—Martians for Justice, Moonlings for Justice, and Venusians for Justice”—are all run by an administrator with an IP address that’s outside the United States. So they’re all connected, even though they’re pretending to not be connected. That’s a very different problem than me sitting in my office in Menlo Park [Calif.] sharing misinformation.

I’m not going to go into lots of technical detail, because this is an area of adversarial nature. The fundamental problem you’re trying to solve is that there’s one entity coordinating the activity of a bunch of things that look like they’re not all one thing. So this is a series of Instagram accounts, or a series of Facebook pages, or a series of WhatsApp accounts, and they’re pretending to be totally different things. We’re looking for signals that these things are related in some way. And we’re looking through the graph [what Facebook calls its map of relationships between users] to understand the properties of this network.

Spectrum: What cutting-edge AI tools and methods have you been working on lately?

Schroepfer: Supervised learning, with humans setting up the instruction process for the AI systems, is amazingly effective. But it has a very obvious flaw: the speed at which you can develop these things is limited by how fast you can curate the data sets. If you’re dealing in a problem domain where things change rapidly, you have to rebuild a new data set and retrain the whole thing.

Self-supervision is inspired by the way people learn, by the way kids explore the world around them. To get computers to do it themselves, we take a bunch of raw data and build a way for the computer to construct its own tests. For language, you scan a bunch of Web pages, and the computer builds a test where it takes a sentence, eliminates one of the words, and figures out how to predict what word belongs there. And because it created the test, it actually knows the answer. I can use as much raw text as I can find and store because it’s processing everything itself and doesn’t require us to sit down and build the information set. In the last two years there has been a revolution in language understanding as a result of AI self-supervised learning.

Spectrum: What else are you excited about?

Schroepfer: What we’ve been working on over the last few years is multilingual understanding. Usually, when I’m trying to figure out, say, whether something is hate speech or not I have to go through the whole process of training the model in every language. I have to do that one time for every language. When you make a post, the first thing we have to figure out is what language your post is in. “Ah, that’s Spanish. So send it to the Spanish hate-speech model.”

We’ve started to build a multilingual model—one box where you can feed in text in 40 different languages and it determines whether it’s hate speech or not. This is way more effective and easier to deploy.

To geek out for a second, just the idea that you can build a model that understands a concept in multiple languages at once is crazy cool. And it not only works for hate speech, it works for a variety of things.

When we started working on this multilingual model years ago, it performed worse than every single individual model. Now, it not only works as well as the English model, but when you get to the languages where you don’t have enough data, it’s so much better. This rapid progress is very exciting.

Spectrum: How do you move new AI tools from your research labs into operational use?

Schroepfer: Engineers trying to make the next breakthrough will often say, “Cool, I’ve got a new thing and it achieved state-of-the-art results on machine translation.” And we say, “Great. How long does it take to run in production?” They say, “Well, it takes 10 seconds for every sentence to run on a CPU.” And we say, “It’ll eat our whole data center if we deploy that.” So we take that state-of-the-art model and we make it 10 or a hundred or a thousand times more efficient, maybe at the cost of a little bit of accuracy. So it’s not as good as the state-of-the-art version, but it’s something we can actually put into our data centers and run in production.

Spectrum: What’s the role of the humans in the loop? Is it true that Facebook currently employs 35,000 moderators?

Schroepfer: Yes. Right now our goal is not to reduce that. Our goal is to do a better job catching bad content. People often think that the end state will be a fully automated system. I don’t see that world coming anytime soon.

As automated systems get more sophisticated, they take more and more of the grunt work away, freeing up the humans to work on the really gnarly stuff where you have to spend an hour researching.

We also use AI to give our human moderators power tools. Say I spot this new meme that is telling everyone to vote on Wednesday rather than Tuesday. I have a tool in front of me that says, “Find variants of that throughout the system. Find every photo with the same text, find every video that mentions this thing and kill it in one shot.” Rather than, I found this one picture, but then a bunch of other people upload that misinformation in different forms.

Another important aspect of AI is that anything I can do to prevent a person from having to look at terrible things is time well spent. Whether it’s a person employed by us as a moderator or a user of our services, looking at these things is a terrible experience. If I can build systems that take the worst of the worst, the really graphic violence, and deal with that in an automated fashion, that’s worth a lot to me. Continue reading

Posted in Human Robots

#437747 High Performance Ornithopter Drone Is ...

The vast majority of drones are rotary-wing systems (like quadrotors), and for good reason: They’re cheap, they’re easy, they scale up and down well, and we’re getting quite good at controlling them, even in very challenging environments. For most applications, though, drones lose out to birds and their flapping wings in almost every way—flapping wings are very efficient, enable astonishing agility, and are much safer, able to make compliant contact with surfaces rather than shredding them like a rotor system does. But flapping wing have their challenges too: Making flapping-wing robots is so much more difficult than just duct taping spinning motors to a frame that, with a few exceptions, we haven’t seen nearly as much improvement as we have in more conventional drones.

In Science Robotics last week, a group of roboticists from Singapore, Australia, China, and Taiwan described a new design for a flapping-wing robot that offers enough thrust and control authority to make stable transitions between aggressive flight modes—like flipping and diving—while also being able to efficiently glide and gently land. While still more complex than a quadrotor in both hardware and software, this ornithopter’s advantages might make it worthwhile.

One reason that making a flapping-wing robot is difficult is because the wings have to move back and forth at high speed while electric motors spin around and around at high speed. This requires a relatively complex transmission system, which (if you don’t do it carefully), leads to weight penalties and a significant loss of efficiency. One particular challenge is that the reciprocating mass of the wings tends to cause the entire robot to flex back and forth, which alternately binds and disengages elements in the transmission system.

The researchers’ new ornithopter design mitigates the flexing problem using hinges and bearings in pairs. Elastic elements also help improve efficiency, and the ornithopter is in fact more efficient with its flapping wings than it would be with a rotary propeller-based propulsion system. Its thrust exceeds its 26-gram mass by 40 percent, which is where much of the aerobatic capability comes from. And one of the most surprising findings of this paper was that flapping-wing robots can actually be more efficient than propeller-based aircraft.

One of the most surprising findings of this paper was that flapping-wing robots can actually be more efficient than propeller-based aircraft

It’s not just thrust that’s a challenge for ornithopters: Control is much more complex as well. Like birds, ornithopters have tails, but unlike birds, they have to rely almost entirely on tail control authority, not having that bird-level of control over fine wing movements. To make an acrobatic level of control possible, the tail control surfaces on this ornithopter are huge—the tail plane area is 35 percent of the wing area. The wings can also provide some assistance in specific circumstances, as by combining tail control inputs with a deliberate stall of the things to allow the ornithopter to execute rapid flips.

With the ability to take off, hover, glide, land softly, maneuver acrobatically, fly quietly, and interact with its environment in a way that’s not (immediately) catastrophic, flapping-wing drones easily offer enough advantages to keep them interesting. Now that ornithopters been shown to be even more efficient than rotorcraft, the researchers plan to focus on autonomy with the goal of moving their robot toward real-world usefulness.

“Efficient flapping wing drone arrests high-speed flight using post-stall soaring,” by Yao-Wei Chin, Jia Ming Kok, Yong-Qiang Zhu, Woei-Leong Chan, Javaan S. Chahl, Boo Cheong Khoo, and Gih-Keong Lau from from Nanyang Technological University in Singapore, National University of Singapore, Defence Science and Technology Group in Canberra, Australia, Qingdao University of Technology in Shandong, China, University of South Australia in Mawson Lakes, and National Chiao Tung University in Hsinchu, Taiwan, was published in Science Robotics. Continue reading

Posted in Human Robots

#437673 Can AI and Automation Deliver a COVID-19 ...

Illustration: Marysia Machulska

Within moments of meeting each other at a conference last year, Nathan Collins and Yann Gaston-Mathé began devising a plan to work together. Gaston-Mathé runs a startup that applies automated software to the design of new drug candidates. Collins leads a team that uses an automated chemistry platform to synthesize new drug candidates.

“There was an obvious synergy between their technology and ours,” recalls Gaston-Mathé, CEO and cofounder of Paris-based Iktos.

In late 2019, the pair launched a project to create a brand-new antiviral drug that would block a specific protein exploited by influenza viruses. Then the COVID-19 pandemic erupted across the world stage, and Gaston-Mathé and Collins learned that the viral culprit, SARS-CoV-2, relied on a protein that was 97 percent similar to their influenza protein. The partners pivoted.

Their companies are just two of hundreds of biotech firms eager to overhaul the drug-discovery process, often with the aid of artificial intelligence (AI) tools. The first set of antiviral drugs to treat COVID-19 will likely come from sifting through existing drugs. Remdesivir, for example, was originally developed to treat Ebola, and it has been shown to speed the recovery of hospitalized COVID-19 patients. But a drug made for one condition often has side effects and limited potency when applied to another. If researchers can produce an ­antiviral that specifically targets SARS-CoV-2, the drug would likely be safer and more effective than a repurposed drug.

There’s one big problem: Traditional drug discovery is far too slow to react to a pandemic. Designing a drug from scratch typically takes three to five years—and that’s before human clinical trials. “Our goal, with the combination of AI and automation, is to reduce that down to six months or less,” says Collins, who is chief strategy officer at SRI Biosciences, a division of the Silicon Valley research nonprofit SRI International. “We want to get this to be very, very fast.”

That sentiment is shared by small biotech firms and big pharmaceutical companies alike, many of which are now ramping up automated technologies backed by supercomputing power to predict, design, and test new antivirals—for this pandemic as well as the next—with unprecedented speed and scope.

“The entire industry is embracing these tools,” says Kara Carter, president of the International Society for Antiviral Research and executive vice president of infectious disease at Evotec, a drug-discovery company in Hamburg. “Not only do we need [new antivirals] to treat the SARS-CoV-2 infection in the population, which is probably here to stay, but we’ll also need them to treat future agents that arrive.”

There are currentlyabout 200 known viruses that infect humans. Although viruses represent less than 14 percent of all known human pathogens, they make up two-thirds of all new human pathogens discovered since 1980.

Antiviral drugs are fundamentally different from vaccines, which teach a person’s immune system to mount a defense against a viral invader, and antibody treatments, which enhance the body’s immune response. By contrast, anti­virals are chemical compounds that directly block a virus after a person has become infected. They do this by binding to specific proteins and preventing them from functioning, so that the virus cannot copy itself or enter or exit a cell.

The SARS-CoV-2 virus has an estimated 25 to 29 proteins, but not all of them are suitable drug targets. Researchers are investigating, among other targets, the virus’s exterior spike protein, which binds to a receptor on a human cell; two scissorlike enzymes, called proteases, that cut up long strings of viral proteins into functional pieces inside the cell; and a polymerase complex that makes the cell churn out copies of the virus’s genetic material, in the form of single-stranded RNA.

But it’s not enough for a drug candidate to simply attach to a target protein. Chemists also consider how tightly the compound binds to its target, whether it binds to other things as well, how quickly it metabolizes in the body, and so on. A drug candidate may have 10 to 20 such objectives. “Very often those objectives can appear to be anticorrelated or contradictory with each other,” says Gaston-Mathé.

Compared with antibiotics, antiviral drug discovery has proceeded at a snail’s pace. Scientists advanced from isolating the first antibacterial molecules in 1910 to developing an arsenal of powerful antibiotics by 1944. By contrast, it took until 1951 for researchers to be able to routinely grow large amounts of virus particles in cells in a dish, a breakthrough that earned the inventors a Nobel Prize in Medicine in 1954.

And the lag between the discovery of a virus and the creation of a treatment can be heartbreaking. According to the World Health Organization, 71 million people worldwide have chronic hepatitis C, a major cause of liver cancer. The virus that causes the infection was discovered in 1989, but effective antiviral drugs didn’t hit the market until 2014.

While many antibiotics work on a range of microbes, most antivirals are highly specific to a single virus—what those in the business call “one bug, one drug.” It takes a detailed understanding of a virus to develop an antiviral against it, says Che Colpitts, a virologist at Queen’s University, in Canada, who works on antivirals against RNA viruses. “When a new virus emerges, like SARS-CoV-2, we’re at a big disadvantage.”

Making drugs to stop viruses is hard for three main reasons. First, viruses are the Spartans of the pathogen world: They’re frugal, brutal, and expert at evading the human immune system. About 20 to 250 nanometers in diameter, viruses rely on just a few parts to operate, hijacking host cells to reproduce and often destroying those cells upon departure. They employ tricks to camouflage their presence from the host’s immune system, including preventing infected cells from sending out molecular distress beacons. “Viruses are really small, so they only have a few components, so there’s not that many drug targets available to start with,” says Colpitts.

Second, viruses replicate quickly, typically doubling in number in hours or days. This constant copying of their genetic material enables viruses to evolve quickly, producing mutations able to sidestep drug effects. The virus that causes AIDS soon develops resistance when exposed to a single drug. That’s why a cocktail of antiviral drugs is used to treat HIV infection.

Finally, unlike bacteria, which can exist independently outside human cells, viruses invade human cells to propagate, so any drug designed to eliminate a virus needs to spare the host cell. A drug that fails to distinguish between a virus and a cell can cause serious side effects. “Discriminating between the two is really quite difficult,” says Evotec’s Carter, who has worked in antiviral drug discovery for over three decades.

And then there’s the money barrier. Developing antivirals is rarely profitable. Health-policy researchers at the London School of Economics recently estimated that the average cost of developing a new drug is US $1 billion, and up to $2.8 billion for cancer and other specialty drugs. Because antivirals are usually taken for only short periods of time or during short outbreaks of disease, companies rarely recoup what they spent developing the drug, much less turn a profit, says Carter.

To change the status quo, drug discovery needs fresh approaches that leverage new technologies, rather than incremental improvements, says Christian Tidona, managing director of BioMed X, an independent research institute in Heidelberg, Germany. “We need breakthroughs.”

Putting Drug Development on Autopilot
Earlier this year, SRI Biosciences and Iktos began collaborating on a way to use artificial intelligence and automated chemistry to rapidly identify new drugs to target the COVID-19 virus. Within four months, they had designed and synthesized a first round of antiviral candidates. Here’s how they’re doing it.

1/5

STEP 1: Iktos’s AI platform uses deep-learning algorithms in an iterative process to come up with new molecular structures likely to bind to and disable a specific coronavirus protein. Illustrations: Chris Philpot

2/5

STEP 2: SRI Biosciences’s SynFini system is a three-part automated chemistry suite for producing new compounds. Starting with a target compound from Iktos, SynRoute uses machine learning to analyze and optimize routes for creating that compound, with results in about 10 seconds. It prioritizes routes based on cost, likelihood of success, and ease of implementation.

3/5

STEP 3: SynJet, an automated inkjet printer platform, tests the routes by printing out tiny quantities of chemical ingredients to see how they react. If the right compound is produced, the platform tests it.

4/5

STEP 4: AutoSyn, an automated tabletop chemical plant, synthesizes milligrams to grams of the desired compound for further testing. Computer-selected “maps” dictate paths through the plant’s modular components.

5/5

STEP 5: The most promising compounds are tested against live virus samples.

Previous
Next

Iktos’s AI platform was created by a medicinal chemist and an AI expert. To tackle SARS-CoV-2, the company used generative models—deep-learning algorithms that generate new data—to “imagine” molecular structures with a good chance of disabling a key coronavirus protein.

For a new drug target, the software proposes and evaluates roughly 1 million compounds, says Gaston-Mathé. It’s an iterative process: At each step, the system generates 100 virtual compounds, which are tested in silico with predictive models to see how closely they meet the objectives. The test results are then used to design the next batch of compounds. “It’s like we have a very, very fast chemist who is designing compounds, testing compounds, getting back the data, then designing another batch of compounds,” he says.

The computer isn’t as smart as a human chemist, Gaston-Mathé notes, but it’s much faster, so it can explore far more of what people in the field call “chemical space”—the set of all possible organic compounds. Unexplored chemical space is huge: Biochemists estimate that there are at least 1063 possible druglike molecules, and that 99.9 percent of all possible small molecules or compounds have never been synthesized.

Still, designing a chemical compound isn’t the hardest part of creating a new drug. After a drug candidate is designed, it must be synthesized, and the highly manual process for synthesizing a new chemical hasn’t changed much in 200 years. It can take days to plan a synthesis process and then months to years to optimize it for manufacture.

That’s why Gaston-Mathé was eager to send Iktos’s AI-generated designs to Collins’s team at SRI Biosciences. With $13.8 million from the Defense Advanced Research Projects Agency, SRI Biosciences spent the last four years automating the synthesis process. The company’s automated suite of three technologies, called SynFini, can produce new chemical compounds in just hours or days, says Collins.

First, machine-learning software devises possible routes for making a desired molecule. Next, an inkjet printer platform tests the routes by printing out and mixing tiny quantities of chemical ingredients to see how they react with one another; if the right compound is produced, the platform runs tests on it. Finally, a tabletop chemical plant synthesizes milligrams to grams of the desired compound.

Less than four months after Iktos and SRI Biosciences announced their collaboration, they had designed and synthesized a first round of antiviral candidates for SARS-CoV-2. Now they’re testing how well the compounds work on actual samples of the virus.

Out of 10
63 possible druglike molecules, 99.9 percent have never been synthesized.

Theirs isn’t the only collaborationapplying new tools to drug discovery. In late March, Alex Zhavoronkov, CEO of Hong Kong–based Insilico Medicine, came across a YouTube video showing three virtual-reality avatars positioning colorful, sticklike fragments in the side of a bulbous blue protein. The three researchers were using VR to explore how compounds might bind to a SARS-CoV-2 enzyme. Zhavoronkov contacted the startup that created the simulation—Nanome, in San Diego—and invited it to examine Insilico’s ­AI-generated molecules in virtual reality.

Insilico runs an AI platform that uses biological data to train deep-learning algorithms, then uses those algorithms to identify molecules with druglike features that will likely bind to a protein target. A four-day training sprint in late January yielded 100 molecules that appear to bind to an important SARS-CoV-2 protease. The company recently began synthesizing some of those molecules for laboratory testing.

Nanome’s VR software, meanwhile, allows researchers to import a molecular structure, then view and manipulate it on the scale of individual atoms. Like human chess players who use computer programs to explore potential moves, chemists can use VR to predict how to make molecules more druglike, says Nanome CEO Steve McCloskey. “The tighter the interface between the human and the computer, the more information goes both ways,” he says.

Zhavoronkov sent data about several of Insilico’s compounds to Nanome, which re-created them in VR. Nanome’s chemist demonstrated chemical tweaks to potentially improve each compound. “It was a very good experience,” says Zhavoronkov.

Meanwhile, in March, Takeda Pharmaceutical Co., of Japan, invited Schrödinger, a New York–based company that develops chemical-simulation software, to join an alliance working on antivirals. Schrödinger’s AI focuses on the physics of how proteins interact with small molecules and one another.

The software sifts through billions of molecules per week to predict a compound’s properties, and it optimizes for multiple desired properties simultaneously, says Karen Akinsanya, chief biomedical scientist and head of discovery R&D at Schrödinger. “There’s a huge sense of urgency here to come up with a potent molecule, but also to come up with molecules that are going to be well tolerated” by the body, she says. Drug developers are seeking compounds that can be broadly used and easily administered, such as an oral drug rather than an intravenous drug, she adds.

Schrödinger evaluated four protein targets and performed virtual screens for two of them, a computing-intensive process. In June, Google Cloud donated the equivalent of 16 million hours of Nvidia GPU time for the company’s calculations. Next, the alliance’s drug companies will synthesize and test the most promising compounds identified by the virtual screens.

Other companies, including Amazon Web Services, IBM, and Intel, as well as several U.S. national labs are also donating time and resources to the Covid-19 High Performance Computing Consortium. The consortium is supporting 87 projects, which now have access to 6.8 million CPU cores, 50,000 GPUs, and 600 petaflops of computational resources.

While advanced technologies could transform early drug discovery, any new drug candidate still has a long road after that. It must be tested in animals, manufactured in large batches for clinical trials, then tested in a series of trials that, for antivirals, lasts an average of seven years.

In May, the BioMed X Institute in Germany launched a five-year project to build a Rapid Antiviral Response Platform, which would speed drug discovery all the way through manufacturing for clinical trials. The €40 million ($47 million) project, backed by drug companies, will identify ­outside-the-box proposals from young scientists, then provide space and funding to develop their ideas.

“We’ll focus on technologies that allow us to go from identification of a new virus to 10,000 doses of a novel potential therapeutic ready for trials in less than six months,” says BioMed X’s Tidona, who leads the project.

While a vaccine will likely arrive long before a bespoke antiviral does, experts expect COVID-19 to be with us for a long time, so the effort to develop a direct-acting, potent antiviral continues. Plus, having new antivirals—and tools to rapidly create more—can only help us prepare for the next pandemic, whether it comes next month or in another 102 years.

“We’ve got to start thinking differently about how to be more responsive to these kinds of threats,” says Collins. “It’s pushing us out of our comfort zones.”

This article appears in the October 2020 print issue as “Automating Antivirals.” Continue reading

Posted in Human Robots