Tag Archives: discussion

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

Posted on November 29, 2020 by Android

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

#437707 Video Friday: This Robot Will Restock ...

Posted on November 25, 2020 by Android

Video Friday is your weekly selection of awesome robotics videos, collected by your Automaton bloggers. We’ll also be posting a weekly calendar of upcoming robotics events for the next few months; here's what we have so far (send us your events!):

CLAWAR 2020 – August 24-26, 2020 – [Online Conference]
ICUAS 2020 – September 1-4, 2020 – Athens, Greece
ICRES 2020 – September 28-29, 2020 – Taipei, Taiwan
AUVSI EXPONENTIAL 2020 – October 5-8, 2020 – [Online Conference]
IROS 2020 – October 25-29, 2020 – Las Vegas, Nev., USA
CYBATHLON 2020 – November 13-14, 2020 – [Online Event]
ICSR 2020 – November 14-16, 2020 – Golden, Colo., USA
Let us know if you have suggestions for next week, and enjoy today's videos.

Tokyo startup Telexistence has recently unveiled a new robot called the Model-T, an advanced teleoperated humanoid that can use tools and grasp a wide range of objects. Japanese convenience store chain FamilyMart plans to test the Model-T to restock shelves in up to 20 stores by 2022. In the trial, a human “pilot” will operate the robot remotely, handling items like beverage bottles, rice balls, sandwiches, and bento boxes.

With Model-T and AWP, FamilyMart and TX aim to realize a completely new store operation by remoteizing and automating the merchandise restocking work, which requires a large number of labor-hours. As a result, stores can operate with less number of workers and enable them to recruit employees regardless of the store’s physical location.

[ Telexistence ]

Quadruped dance-off should be a new robotics competition at IROS or ICRA.

I dunno though, that moonwalk might keep Spot in the lead…

[ Unitree ]

Through a hybrid of simulation and real-life training, this air muscle robot is learning to play table tennis.

Table tennis requires to execute fast and precise motions. To gain precision it is necessary to explore in this high-speed regimes, however, exploration can be safety-critical at the same time. The combination of RL and muscular soft robots allows to close this gap. While robots actuated by pneumatic artificial muscles generate high forces that are required for e.g. smashing, they also offer safe execution of explosive motions due to antagonistic actuation.

To enable practical training without real balls, we introduce Hybrid Sim and Real Training (HYSR) that replays prerecorded real balls in simulation while executing actions on the real system. In this manner, RL can learn the challenging motor control of the PAM-driven robot while executing ~15000 hitting motions.

[ Max Planck Institute ]

Thanks Dieter!

Anthony Cowley wrote in to share his recent thesis work on UPSLAM, a fast and lightweight SLAM technique that records data in panoramic depth images (just PNGs) that are easy to visualize and even easier to share between robots, even on low-bandwidth networks.

[ UPenn ]

Thanks Anthony!

GITAI’s G1 is the space dedicated general-purpose robot. G1 robot will enable automation of various tasks internally & externally on space stations and for lunar base development.

[ Gitai ]

The University of Michigan has a fancy new treadmill that’s built right into the floor, which proves to be a bit much for Mini Cheetah.

But Cassie Blue won’t get stuck on no treadmill! She goes for a 0.3 mile walk across campus, which ends when a certain someone ran the gantry into Cassie Blue’s foot.

[ Michigan Robotics ]

Some serious quadruped research going on at UT Austin Human Centered Robotics Lab.

[ HCRL ]

Will Burrard-Lucas has spent lockdown upgrading his slightly indestructible BeetleCam wildlife photographing robot.

[ Will Burrard-Lucas ]

Teleoperated surgical robots are becoming commonplace in operating rooms, but many are massive (sometimes taking up an entire room) and are difficult to manipulate, especially if a complication arises and the robot needs to removed from the patient. A new collaboration between the Wyss Institute, Harvard University, and Sony Corporation has created the mini-RCM, a surgical robot the size of a tennis ball that weighs as much as a penny, and performed significantly better than manually operated tools in delicate mock-surgical procedures. Importantly, its small size means it is more comparable to the human tissues and structures on which it operates, and it can easily be removed by hand if needed.

[ Harvard Wyss ]

Yaskawa appears to be working on a robot that can scan you with a temperature gun and then jam a mask on your face?

[ Motoman ]

Maybe we should just not have people working in mines anymore, how about that?

[ Exyn ]

Many current human-robot interactive systems tend to use accurate and fast – but also costly – actuators and tracking systems to establish working prototypes that are safe to use and deploy for user studies. This paper presents an embedded framework to build a desktop space for human-robot interaction, using an open-source robot arm, as well as two RGB cameras connected to a Raspberry Pi-based controller that allow a fast yet low-cost object tracking and manipulation in 3D. We show in our evaluations that this facilitates prototyping a number of systems in which user and robot arm can commonly interact with physical objects.

[ Paper ]

IBM Research is proud to host professor Yoshua Bengio — one of the world’s leading experts in AI — in a discussion of how AI can contribute to the fight against COVID-19.

[ IBM Research ]

Ira Pastor, ideaXme life sciences ambassador interviews Professor Dr. Hiroshi Ishiguro, the Director of the Intelligent Robotics Laboratory, of the Department of Systems Innovation, in the Graduate School of Engineering Science, at Osaka University, Japan.

[ ideaXme ]

A CVPR talk from Stanford’s Chelsea Finn on “Generalization in Visuomotor Learning.”

[ Stanford ] Continue reading →

Posted in Human Robots

#437564 How We Won the DARPA SubT Challenge: ...

Posted on November 18, 2020 by Android

This is a guest post. The views expressed here are those of the authors and do not necessarily represent positions of IEEE or its organizational units.​

“Do you smell smoke?” It was three days before the qualification deadline for the Virtual Tunnel Circuit of the DARPA Subterranean Challenge Virtual Track, and our team was barrelling through last-minute updates to our robot controllers in a small conference room at the Michigan Tech Research Institute (MTRI) offices in Ann Arbor, Mich. That’s when we noticed the smell. We’d assumed that one of the benefits of entering a virtual disaster competition was that we wouldn’t be exposed to any actual disasters, but equipment in the basement of the building MTRI shares had started to smoke. We evacuated. The fire department showed up. And as soon as we could, the team went back into the building, hunkered down, and tried to make up for the unexpected loss of several critical hours.

Team BARCS joins the SubT Virtual Track
The smoke incident happened more than a year after we first learned of the DARPA Subterranean Challenge. DARPA announced SubT early in 2018, and at that time, we were interested in building internal collaborations on multi-agent autonomy problems, and SubT seemed like the perfect opportunity. Though a few of us had backgrounds in robotics, the majority of our team was new to the field. We knew that submitting a proposal as a largely non-traditional robotics team from an organization not known for research in robotics was a risk. However, the Virtual Track gave us the opportunity to focus on autonomy and multi-agent teaming strategies, areas requiring skill in asynchronous computing and sensor data processing that are strengths of our Institute. The prevalence of open source code, small inexpensive platforms, and customizable sensors has provided the opportunity for experts in fields other than robotics to apply novel approaches to robotics problems. This is precisely what makes the Virtual Track of SubT appealing to us, and since starting SubT, autonomy has developed into a significant research thrust for our Institute. Plus, robots are fun!

After many hours of research, discussion, and collaboration, we submitted our proposal early in 2018. And several months later, we found out that we had won a contract and became a funded team (Team BARCS) in the SubT Virtual Track. Now we needed to actually make our strategy work for the first SubT Tunnel Circuit competition, taking place in August of 2019.

Building a team of virtual robots
A natural approach to robotics competitions like SubT is to start with the question of “what can X-type robot do” and then build a team and strategy around individual capabilities. A particular challenge for the SubT Virtual Track is that we can’t design our own systems; instead, we have to choose from a predefined set of simulated robots and sensors that DARPA provides, based on the real robots used by Systems Track teams. Our approach is to look at what a team of robots can do together, determining experimentally what the best team configuration is for each environment. By the final competition, ideally we will be demonstrating the value of combining platforms across multiple Systems Track teams into a single Virtual Track team. Each of the robot configurations in the competition has an associated cost, and team size is constrained by a total cost. This provides another impetus for limiting dependence on complex sensor packages, though our ranging preference is 3D lidar, which is the most expensive sensor!

Image: Michigan Tech Research Institute

The teams can rely on realistic physics and sensors but they start off with no maps of any kind, so the focus is on developing autonomous exploratory behavior, navigation methods, and object recognition for their simulated robots.

One of the frequent questions we receive about the Virtual Track is if it’s like a video game. While it may look similar on the surface, everything under the hood in a video game is designed to service the game narrative and play experience, not require novel research in AI and autonomy. The purpose of simulations, on the other hand, is to include full physics and sensor models (including noise and errors) to provide a testbed for prototyping and developing solutions to those real-world challenges. We are starting with realistic physics and sensors but no maps of any kind, so the focus is on developing autonomous exploratory behavior, navigation methods, and object recognition for our simulated robots.

Though the simulation is more like real life than a video game, it is not real life. Due to occasional software bugs, there are still non-physical events, like the robots falling through an invisible hole in the world or driving through a rock instead of over it or flipping head over heels when driving over a tiny lip between world tiles. These glitches, while sometimes frustrating, still allow the SubT Virtual platform to be realistic enough to support rapid prototyping of controller modules that will transition straightforwardly onto hardware, closing the loop between simulation and real-world robots.

Full autonomy for DARPA-hard scenarios
The Virtual Track requirement that the robotic agents be fully autonomous, rather than have a human supervisor, is a significant distinction between the Systems and Virtual Tracks of SubT. Our solutions must be hardened against software faults caused by things like missing and bad data since our robots can’t turn to us for help. In order for a team of robots to complete this objective reliably with no human-in-the-loop, all of the internal systems, from perception to navigation to control to actuation to communications, must be able to autonomously identify and manage faults and failures anywhere in the control chain.

The communications limitations in subterranean environments (both real and virtual) mean that we need to keep the amount of information shared between robots low, while making the usability of that information for joint decision-making high. This goal has guided much of our design for autonomous navigation and joint search strategy for our team. For example, instead of sharing the full SLAM map of the environment, our agents only share a simplified graphical representation of the space, along with data about frontiers it has not yet explored, and are able to merge its information with the graphs generated by other agents. The merged graph can then be used for planning and navigation without having full knowledge of the detailed 3D map.

The Virtual Track requires that the robotic agents be fully autonomous. With no human-in-the-loop, all of the internal systems, from perception to navigation to control to actuation to communications, must be able to identify and manage faults and failures anywhere in the control chain.

Since the objective of the SubT program is to advance the state-of-the-art in rapid autonomous exploration and mapping of subterranean environments by robots, our first software design choices focused on the mapping task. The SubT virtual environments are sufficiently rich as to provide interesting problems in building so-called costmaps that accurately separate obstructions that are traversable (like ramps) from legitimately impassible obstructions. An extra complication we discovered in the first course, which took place in mining tunnels, was that the angle of the lowest beam of the lidar was parallel to the down ramps in the tunnel environment, so they could not “see” the ground (or sometimes even obstructions on the ramp) until they got close enough to the lip of the ramp to receive lidar reflections off the bottom of the ramp. In this case, we had to not only change the costmap to convince the robot that there was safe ground to reach over the lip of the ramp, but also had to change the path planner to get the robot to proceed with caution onto the top of the ramp in case there were previously unseen obstructions on the ramp.

In addition to navigation in the costmaps, the robot must be able to generate its own goals to navigate to. This is what produces exploratory behavior when there is no map to start with. SLAM is used to generate a detailed map of the environment explored by a single robot—the space it has probed with its sensors. From the sensor data, we are able to extract information about the interior space of the environment while looking for holes in the data, to determine things like whether the current tunnel continues or ends, or how many tunnels meet at an intersection. Once we have some understanding of the interior space, we can place navigation goals in that space. These goals naturally update as the robot traverses the tunnel, allowing the entire space to be explored.

Sending our robots into the virtual unknown
The solutions for the Virtual Track competitions are tested by DARPA in multiple sequestered runs across many environments for each Circuit in the month prior to the Systems Track competition. We must wait until the joint award ceremony at the conclusion of the Systems Track to find out the results, and we are completely in the dark about placings before the awards are announced. It’s nerve-wracking! The challenges of the worlds used in the Circuit events are also hand-designed, so features of the worlds we use for development could be combined in ways we have not anticipated—it’s always interesting to see what features were prioritized after the event. We test everything in our controllers well enough to feel confident that we at least are submitting something reasonably stable and broadly capable, and once the solution is in, we can’t really do anything other than “let go” and get back to work on the next phase of development. Maybe it’s somewhat like sending your kid to college: “we did our best to prepare you for this world, little bots. Go do good.”

Image: Michigan Tech Research Institute

The first SubT competition was the Tunnel Circuit, featuring a labyrinthine environment that simulated human-engineered tunnels, including hazards such as vertical shafts and rubble.

The first competition was the Tunnel Circuit, in October 2019. This environment models human-engineered tunnels. Two substantial challenges in this environment were vertical shafts and rubble. Our team accrued 21 points over 15 competition runs in five separate tunnel environments for a second place finish, behind Team Coordinated Robotics.

The next phase of the SubT virtual competition was the Urban Circuit. Much of the difference between our Tunnel and Urban Circuit results came down to thorough testing to identify failure modes and implementations of checks and data filtering for fault tolerance. For example, in the SLAM nodes run by a single robot, the coordinates of the most recent sensor data are changed multiple times during processing and integration into the current global 3D map of the “visited” environment stored by that robot. If there is lag in IMU or clock data, the observation may be temporarily registered at a default location that is very far from the actual position. Since most of our decision processes for exploration are downstream from SLAM, this can cause faulty or impossible goals to be generated, and the robots then spend inordinate amounts of time trying to drive through walls. We updated our method to add a check to see if the new map position has jumped a far distance from the prior map position, and if so, we threw that data out.

Image: Michigan Tech Research Institute

In open spaces like the rooms in the Urban circuit, we adjusted our approach to exploration through graph generation to allow the robots to accurately identify viable routes while helping to prevent forays off platform edges.

Our approach to exploration through graph generation based on identification of interior spaces allowed us to thoroughly explore the centers of rooms, although we did have to make some changes from the Tunnel circuit to achieve that. In the Tunnel circuit, we used a simplified graph of the environment based on landmarks like intersections. The advantage of this approach is that it is straightforward for two robots to compare how the graphs of the space they explored individually overlap. In open spaces like the rooms in the Urban circuit, we chose to instead use a more complex, less directly comparable graph structure based on the individual robot’s trajectory. This allowed the robots to accurately identify viable routes between features like subway station platforms and subway tracks, as well as to build up the navigation space for room interiors, while helping to prevent forays off the platform edges. Frontier information is also integrated into the graph, providing a uniform data structure for both goal selection and route planning.

The results are in!
The award ceremony for the Urban Circuit was held concurrently with the Systems Track competition awards this past February in Washington State. We sent a team representative to participate in the Technical Interchange Meeting and present the approach for our team, and the rest of us followed along from our office space on the DARPAtv live stream. While we were confident in our solution, we had also been tracking the online leaderboard and knew our competitors were going to be submitting strong solutions. Since the competition environments are hand-designed, there are always novel challenges that could be presented in these environments as well. We knew we would put up a good fight, but it was very exciting to see BARCS appear in first place!

Any time we implement a new module in our control system, there is a lot of parameter tuning that has to happen to produce reliably good autonomous behavior. In the Urban Circuit, we did not sufficiently test some parameter values in our exploration modules. The effect of this was that the robots only chose to go down small hallways after they explored everything else in their environment, which meant very often they ran out of time and missed a lot of small rooms. This may be the biggest source of lost points for us in the Urban Circuit. One of our major plans going forward from the Urban Circuit is to integrate more sophisticated node selection methods, which can help our robots more intelligently prioritize which frontier nodes to visit. By going through all three Circuit challenges, we will learn how to appropriately add weights to the frontiers based on features of the individual environments. For the Final Challenge, when all three Circuit environments will be combined into large systems, we plan to implement adaptive controllers that will identify their environments and use the appropriate optimized parameters for that environment. In this way, we expect our agents to be able to (for example) prioritize hallways and other small spaces in Urban environments, and perhaps prioritize large openings over small in the Cave environments, if the small openings end up being treacherous overall.

Next for our team: Cave Circuit
Coming up next for Team BARCS is the Virtual Cave Circuit. We are in the middle of testing our hypothesis that our controller will transition from UGVs to UAVs and developing strategies for refining our solution to handle Cave Circuit environmental hazards. The UAVs have a shorter battery life than the UGVs, so executing a joint exploration strategy will also be a high priority for this event, as will completing our work on graph sharing and merging, which will give our robot teams more sophisticated options for navigation and teamwork. We’re reaching a threshold in development where we can start increasing the “smarts” of the robots, which we anticipate will be critical for the final competition, where all of the challenges of SubT will be combined to push the limits of innovation. The Cave Circuit will also have new environmental challenges to tackle: dynamic features such as rock falls have been added, which will block previously accessible passages in the cave environment. We think our controllers are well-poised to handle this new challenge, and we’re eager to find out if that’s the case.

As of now, the biggest worries for us are time and team composition. The Cave Circuit deadline has been postponed to October 15 due to COVID-19 delays, with the award ceremony in mid-November, but there have also been several very compelling additions to the testbed that we would like to experiment with before submission, including droppable networking ‘breadcrumbs’ and new simulated platforms. There are design trade-offs when balancing general versus specialist approaches to the controllers for these robots—since we are adding UAVs to our team for the first time, there are new decisions that will have to be made. For example, the UAVs can ascend into vertical spaces, but only have a battery life of 20 minutes. The UGVs by contrast have 90 minute battery life. One of our strategies is to do an early return to base with one or more agents to buy down risk on making any artifact reports at all for the run, hedging against our other robots not making it back in time, a lesson learned from the Tunnel Circuit. Should a UAV take on this role, or is it better to have them explore deeper into the environment and instead report their artifacts to a UGV or network node, which comes with its own risks? Testing and experimentation to determine the best options takes time, which is always a worry when preparing for a competition! We also anticipate new competitors and stiffer competition all around.

Image: Michigan Tech Research Institute

Team BARCS has now a year to prepare for the final DARPA SubT Challenge event, expected to take place in late 2021.

Going forward from the Cave Circuit, we will have a year to prepare for the final DARPA SubT Challenge event, expected to take place in late 2021. What we are most excited about is increasing the level of intelligence of the agents in their teamwork and joint exploration of the environment. Since we will have (hopefully) built up robust approaches to handling each of the specific types of environments in the Tunnel, Urban, and Cave circuits, we will be aiming to push the limits on collaboration and efficiency among the agents in our team. We view this as a central research contribution of the Virtual Track to the Subterranean Challenge because intelligent, adaptive, multi-robot collaboration is an upcoming stage of development for integration of robots into our lives.

The Subterranean Challenge Virtual Track gives us a bridge for transitioning our more abstract research ideas and algorithms relevant to this degree of autonomy and collaboration onto physical systems, and exploring the tangible outcomes of implementing our work in the real world. And the next time there’s an incident in the basement of our building, the robots (and humans) of Team BARCS will be ready to respond.

Richard Chase, Ph.D., P.E., is a research scientist at Michigan Tech Research Institute (MTRI) and has 20 years of experience developing robotics and cyber physical systems in areas from remote sensing to autonomous vehicles. At MTRI, he works on a variety of topics such as swarm autonomy, human-swarm teaming, and autonomous vehicles. His research interests are the intersection of design, robotics, and embedded systems.

Sarah Kitchen is a Ph.D. mathematician working as a research scientist and an AI/Robotics focus area leader at MTRI. Her research interests include intelligent autonomous agents and multi-agent collaborative teams, as well as applications of autonomous robots to sensing systems.

This material is based upon work supported by the Defense Advanced Research Projects Agency (DARPA) under Contract No. HR001118C0124 and is released under Distribution Statement (Approved for Public Release, Distribution Unlimited). Any opinions, findings and conclusions or recommendations expressed in this material are those of the authors and do not necessarily reflect the views of DARPA. Continue reading →

Posted in Human Robots

#437562 Video Friday: Aquanaut Robot Takes to ...

Posted on November 18, 2020 by Android

Video Friday is your weekly selection of awesome robotics videos, collected by your Automaton bloggers. We’ll also be posting a weekly calendar of upcoming robotics events for the next few months; here's what we have so far (send us your events!):

IROS 2020 – October 25-25, 2020 – [Online]
ICSR 2020 – November 14-16, 2020 – Golden, Colo., USA
Bay Area Robotics Symposium – November 20, 2020 – [Online]
ACRA 2020 – December 8-10, 2020 – [Online]
Let us know if you have suggestions for next week, and enjoy today's videos.

To prepare the Perseverance rover for its date with Mars, NASA’s Mars 2020 mission team conducted a wide array of tests to help ensure a successful entry, descent and landing at the Red Planet. From parachute verification in the world’s largest wind tunnel, to hazard avoidance practice in Death Valley, California, to wheel drop testing at NASA’s Jet Propulsion Laboratory and much more, every system was put through its paces to get ready for the big day. The Perseverance rover is scheduled to land on Mars on February 18, 2021.

[ JPL ]

Awesome to see Aquanaut—the “underwater transformer” we wrote about last year—take to the ocean!

Also their new website has SHARKS on it.

[ HMI ]

Nature has inspired engineers at UNSW Sydney to develop a soft fabric robotic gripper which behaves like an elephant's trunk to grasp, pick up and release objects without breaking them.

[ UNSW ]

Collaborative robots offer increased interaction capabilities at relatively low cost but, in contrast to their industrial counterparts, they inevitably lack precision. We address this problem by relying on a dual-arm system with laser-based sensing to measure relative poses between objects of interest and compensate for pose errors coming from robot proprioception.

[ Paper ]

Developed by NAVER LABS, with Korea University of Technology & Education (Koreatech), the robot arm now features an added waist, extending the available workspace, as well as a sensor head that can perceive objects. It has also been equipped with a robot hand “BLT Gripper” that can change to various grasping methods.

[ NAVER Labs ]

In case you were still wondering why SoftBank acquired Aldebaran and Boston Dynamics:

[ RobotStart ]

DJI's new Mini 2 drone is here with a commercial so hip it makes my teeth scream.

[ DJI ]

Using simple materials, such as plastic struts and cardboard rolls, the first prototype of the RBO Hand 3 is already capable of grasping a large range of different objects thanks to its opposable thumb.

The RBO Hand 3 performs an edge grasp before handing-over the object to a person. The hand actively exploits constraints in the environment (the tabletop) for grasping the object. Thanks to its compliance, this interaction is safe and robust.

[ TU Berlin ]

Flyability's Elios 2 helped researchers inspect Reactor Five at the Chernobyl nuclear disaster site in order to determine whether any uranium was present. Prior to this mission, Reactor Five had not been investigated since the disaster in April of 1986.

[ Flyability ]

Thanks Zacc!

SOTO 2 is here! Together with our development partners from the industry, we have greatly enhanced the SOTO prototype over the last two years. With the new version of the robot, Industry 4.0 will become a great deal more real: SOTO brings materials to the assembly line, just-in-time and completely autonomously.

[ Magazino ]

A drone that can fly sustainably for long distances over land and water, and can land almost anywhere, will be able to serve a wide range of applications. There are already drones that fly using ‘green’ hydrogen, but they either fly very slowly or cannot land vertically. That’s why researchers at TU Delft, together with the Royal Netherlands Navy and the Netherlands Coastguard, developed a hydrogen-powered drone that is capable of vertical take-off and landing whilst also being able to fly horizontally efficiently for several hours, much like regular aircraft. The drone uses a combination of hydrogen and batteries as its power source.

[ MAVLab ]

The National Nuclear User Facility for Hot Robotics (NNUF-HR) is an EPSRC funded facility to support UK academia and industry to deliver ground-breaking, impactful research in robotics and artificial intelligence for application in extreme and challenging nuclear environments.

[ NNUF ]

At the Karolinska University Laboratory in Sweden, an innovation project based around an ABB collaborative robot has increased efficiency and created a better working environment for lab staff.

[ ABB ]

What I find interesting about DJI's enormous new agricultural drone is that it's got a spinning obstacle detecting sensor that's a radar, not a lidar.

Also worth noting is that it seems to detect the telephone pole, but not the support wire that you can see in the video feed, although the visualization does make it seem like it can spot the power lines above.

[ DJI ]

Josh Pieper has spend the last year building his own quadruped, and you can see what he's been up to in just 12 minutes.

[ mjbots ]

Thanks Josh!

Dr. Ryan Eustice, TRI Senior Vice President of Automated Driving, delivers a keynote speech — “The Road to Vehicle Automation, a Toyota Guardian Approach” — to SPIE's Future Sensing Technologies 2020. During the presentation, Eustice provides his perspective on the current state of automated driving, summarizes TRI's Guardian approach — which amplifies human drivers, rather than replacing them — and summarizes TRI's recent developments in core AD capabilities.

[ TRI ]

Two excellent talks this week from UPenn GRASP Lab, from Ruzena Bajcsy and Vijay Kumar.

A panel discussion on the future of robotics and societal challenges with Dr. Ruzena Bajcsy as a Roboticist and Founder of the GRASP Lab.

In this talk I will describe the role of the White House Office of Science and Technology Policy in supporting science and technology research and education, and the lessons I learned while serving in the office. I will also identify a few opportunities at the intersection of technology and policy and broad societal challenges.

[ UPenn ]

The IROS 2020 “Perception, Learning, and Control for Autonomous Agile Vehicles” workshop is all online—here's the intro, but you can click through for a playlist that includes videos of the entire program, and slides are available as well.

[ NYU ] Continue reading →

Posted in Human Robots

#437301 The Global Work Crisis: Automation, the ...

Posted on August 7, 2020 by Android

The alarm bell rings. You open your eyes, come to your senses, and slide from dream state to consciousness. You hit the snooze button, and eventually crawl out of bed to the start of yet another working day.

This daily narrative is experienced by billions of people all over the world. We work, we eat, we sleep, and we repeat. As our lives pass day by day, the beating drums of the weekly routine take over and years pass until we reach our goal of retirement.

A Crisis of Work
We repeat the routine so that we can pay our bills, set our kids up for success, and provide for our families. And after a while, we start to forget what we would do with our lives if we didn’t have to go back to work.

In the end, we look back at our careers and reflect on what we’ve achieved. It may have been the hundreds of human interactions we’ve had; the thousands of emails read and replied to; the millions of minutes of physical labor—all to keep the global economy ticking along.

According to Gallup’s World Poll, only 15 percent of people worldwide are actually engaged with their jobs. The current state of “work” is not working for most people. In fact, it seems we as a species are trapped by a global work crisis, which condemns people to cast away their time just to get by in their day-to-day lives.

Technologies like artificial intelligence and automation may help relieve the work burdens of millions of people—but to benefit from their impact, we need to start changing our social structures and the way we think about work now.

The Specter of Automation
Automation has been ongoing since the Industrial Revolution. In recent decades it has taken on a more elegant guise, first with physical robots in production plants, and more recently with software automation entering most offices.

The driving goal behind much of this automation has always been productivity and hence, profits: technology that can act as a multiplier on what a single human can achieve in a day is of huge value to any company. Powered by this strong financial incentive, the quest for automation is growing ever more pervasive.

But if automation accelerates or even continues at its current pace and there aren’t strong social safety nets in place to catch the people who are negatively impacted (such as by losing their jobs), there could be a host of knock-on effects, including more concentrated wealth among a shrinking elite, more strain on government social support, an increase in depression and drug dependence, and even violent social unrest.

It seems as though we are rushing headlong into a major crisis, driven by the engine of accelerating automation. But what if instead of automation challenging our fragile status quo, we view it as the solution that can free us from the shackles of the Work Crisis?

The Way Out
In order to undertake this paradigm shift, we need to consider what society could potentially look like, as well as the problems associated with making this change. In the context of these crises, our primary aim should be for a system where people are not obligated to work to generate the means to survive. This removal of work should not threaten access to food, water, shelter, education, healthcare, energy, or human value. In our current system, work is the gatekeeper to these essentials: one can only access these (and even then often in a limited form), if one has a “job” that affords them.

Changing this system is thus a monumental task. This comes with two primary challenges: providing people without jobs with financial security, and ensuring they maintain a sense of their human value and worth. There are several measures that could be implemented to help meet these challenges, each with important steps for society to consider.

Universal basic income (UBI)

UBI is rapidly gaining support, and it would allow people to become shareholders in the fruits of automation, which would then be distributed more broadly.

UBI trials have been conducted in various countries around the world, including Finland, Kenya, and Spain. The findings have generally been positive on the health and well-being of the participants, and showed no evidence that UBI disincentivizes work, a common concern among the idea’s critics. The most recent popular voice for UBI has been that of former US presidential candidate Andrew Yang, who now runs a non-profit called Humanity Forward.

UBI could also remove wasteful bureaucracy in administering welfare payments (since everyone receives the same amount, there’s no need to prevent false claims), and promote the pursuit of projects aligned with peoples’ skill sets and passions, as well as quantifying the value of tasks not recognized by economic measures like Gross Domestic Product (GDP). This includes looking after children and the elderly at home.

How a UBI can be initiated with political will and social backing and paid for by governments has been hotly debated by economists and UBI enthusiasts. Variables like how much the UBI payments should be, whether to implement taxes such as Yang’s proposed valued added tax (VAT), whether to replace existing welfare payments, the impact on inflation, and the impact on “jobs” from people who would otherwise look for work require additional discussion. However, some have predicted the inevitability of UBI as a result of automation.

Universal healthcare

Another major component of any society is the healthcare of its citizens. A move away from work would further require the implementation of a universal healthcare system to decouple healthcare from jobs. Currently in the US, and indeed many other economies, healthcare is tied to employment.

Universal healthcare such as Medicare in Australia is evidence for the adage “prevention is better than cure,” when comparing the cost of healthcare in the US with Australia on a per capita basis. This has already presented itself as an advancement in the way healthcare is considered. There are further benefits of a healthier population, including less time and money spent on “sick-care.” Healthy people are more likely and more able to achieve their full potential.

Reshape the economy away from work-based value

One of the greatest challenges in a departure from work is for people to find value elsewhere in life. Many people view their identities as being inextricably tied to their jobs, and life without a job is therefore a threat to one’s sense of existence. This presents a shift that must be made at both a societal and personal level.

A person can only seek alternate value in life when afforded the time to do so. To this end, we need to start reducing “work-for-a-living” hours towards zero, which is a trend we are already seeing in Europe. This should not come at the cost of reducing wages pro rata, but rather could be complemented by UBI or additional schemes where people receive dividends for work done by automation. This transition makes even more sense when coupled with the idea of deviating from using GDP as a measure of societal growth, and instead adopting a well-being index based on universal human values like health, community, happiness, and peace.

The crux of this issue is in transitioning away from the view that work gives life meaning and life is about using work to survive, towards a view of living a life that itself is fulfilling and meaningful. This speaks directly to notions from Maslow’s hierarchy of needs, where work largely addresses psychological and safety needs such as shelter, food, and financial well-being. More people should have a chance to grow beyond the most basic needs and engage in self-actualization and transcendence.

The question is largely around what would provide people with a sense of value, and the answers would differ as much as people do; self-mastery, building relationships and contributing to community growth, fostering creativity, and even engaging in the enjoyable aspects of existing jobs could all come into play.

Universal education

With a move towards a society that promotes the values of living a good life, the education system would have to evolve as well. Researchers have long argued for a more nimble education system, but universities and even most online courses currently exist for the dominant purpose of ensuring people are adequately skilled to contribute to the economy. These “job factories” only exacerbate the Work Crisis. In fact, the response often given by educational institutions to the challenge posed by automation is to find new ways of upskilling students, such as ensuring they are all able to code. As alluded to earlier, this is a limited and unimaginative solution to the problem we are facing.

Instead, education should be centered on helping people acknowledge the current crisis of work and automation, teach them how to derive value that is decoupled from work, and enable people to embrace progress as we transition to the new economy.

Disrupting the Status Quo
While we seldom stop to think about it, much of the suffering faced by humanity is brought about by the systemic foe that is the Work Crisis. The way we think about work has brought society far and enabled tremendous developments, but at the same time it has failed many people. Now the status quo is threatened by those very developments as we progress to an era where machines are likely to take over many job functions.

This impending paradigm shift could be a threat to the stability of our fragile system, but only if it is not fully anticipated. If we prepare for it appropriately, it could instead be the key not just to our survival, but to a better future for all.

Image Credit: mostafa meraji from Pixabay Continue reading →

Posted in Human Robots