Tag Archives: apply

#437624 AI-Powered Drone Learns Extreme ...

Quadrotors are among the most agile and dynamic machines ever created. In the hands of a skilled human pilot, they can do some astonishing series of maneuvers. And while autonomous flying robots have been getting better at flying dynamically in real-world environments, they still haven’t demonstrated the same level of agility of manually piloted ones.

Now researchers from the Robotics and Perception Group at the University of Zurich and ETH Zurich, in collaboration with Intel, have developed a neural network training method that “enables an autonomous quadrotor to fly extreme acrobatic maneuvers with only onboard sensing and computation.” Extreme.

There are two notable things here: First, the quadrotor can do these extreme acrobatics outdoors without any kind of external camera or motion-tracking system to help it out (all sensing and computing is onboard). Second, all of the AI training is done in simulation, without the need for an additional simulation-to-real-world (what researchers call “sim-to-real”) transfer step. Usually, a sim-to-real transfer step means putting your quadrotor into one of those aforementioned external tracking systems, so that it doesn’t completely bork itself while trying to reconcile the differences between the simulated world and the real world, where, as the researchers wrote in a paper describing their system, “even tiny mistakes can result in catastrophic outcomes.”

To enable “zero-shot” sim-to-real transfer, the neural net training in simulation uses an expert controller that knows exactly what’s going on to teach a “student controller” that has much less perfect knowledge. That is, the simulated sensory input that the student ends up using as it learns to follow the expert has been abstracted to present the kind of imperfect, imprecise data it’s going to encounter in the real world. This can involve things like abstracting away the image part of the simulation until you’d have no way of telling the difference between abstracted simulation and abstracted reality, which is what allows the system to make that sim-to-real leap.

The simulation environment that the researchers used was Gazebo, slightly modified to better simulate quadrotor physics. Meanwhile, over in reality, a custom 1.5-kilogram quadrotor with a 4:1 thrust to weight ratio performed the physical experiments, using only a Nvidia Jetson TX2 computing board and an Intel RealSense T265, a dual fisheye camera module optimized for V-SLAM. To challenge the learning system, it was trained to perform three acrobatic maneuvers plus a combo of all of them:

Image: University of Zurich/ETH Zurich/Intel

Reference trajectories for acrobatic maneuvers. Top row, from left: Power Loop, Barrel Roll, and Matty Flip. Bottom row: Combo.

All of these maneuvers require high accelerations of up to 3 g’s and careful control, and the Matty Flip is particularly challenging, at least for humans, because the whole thing is done while the drone is flying backwards. Still, after just a few hours of training in simulation, the drone was totally real-world competent at these tricks, and could even extrapolate a little bit to perform maneuvers that it was not explicitly trained on, like doing multiple loops in a row. Where humans still have the advantage over drones is (as you might expect since we’re talking about robots) is quickly reacting to novel or unexpected situations. And when you’re doing this sort of thing outdoors, novel and unexpected situations are everywhere, from a gust of wind to a jealous bird.

For more details, we spoke with Antonio Loquercio from the University of Zurich’s Robotics and Perception Group.

IEEE Spectrum: Can you explain how the abstraction layer interfaces with the simulated sensors to enable effective sim-to-real transfer?

Antonio Loquercio: The abstraction layer applies a specific function to the raw sensor information. Exactly the same function is applied to the real and simulated sensors. The result of the function, which is “abstracted sensor measurements,” makes simulated and real observation of the same scene similar. For example, suppose we have a sequence of simulated and real images. We can very easily tell apart the real from the simulated ones given the difference in rendering. But if we apply the abstraction function of “feature tracks,” which are point correspondences in time, it becomes very difficult to tell which are the simulated and real feature tracks, since point correspondences are independent of the rendering. This applies for humans as well as for neural networks: Training policies on raw images gives low sim-to-real transfer (since images are too different between domains), while training on the abstracted images has high transfer abilities.

How useful is visual input from a camera like the Intel RealSense T265 for state estimation during such aggressive maneuvers? Would using an event camera substantially improve state estimation?

Our end-to-end controller does not require a state estimation module. It shares however some components with traditional state estimation pipelines, specifically the feature extractor and the inertial measurement unit (IMU) pre-processing and integration function. The input of the neural networks are feature tracks and integrated IMU measurements. When looking at images with low features (for example when the camera points to the sky), the neural net will mainly rely on IMU. When more features are available, the network uses to correct the accumulated drift from IMU. Overall, we noticed that for very short maneuvers IMU measurements were sufficient for the task. However, for longer ones, visual information was necessary to successfully address the IMU drift and complete the maneuver. Indeed, visual information reduces the odds of a crash by up to 30 percent in the longest maneuvers. We definitely think that event camera can improve even more the current approach since they could provide valuable visual information during high speed.

“The Matty Flip is probably one of the maneuvers that our approach can do very well … It is super challenging for humans, since they don’t see where they’re going and have problems in estimating their speed. For our approach the maneuver is no problem at all, since we can estimate forward velocities as well as backward velocities.”
—Antonio Loquercio, University of Zurich

You describe being able to train on “maneuvers that stretch the abilities of even expert human pilots.” What are some examples of acrobatics that your drones might be able to do that most human pilots would not be capable of?

The Matty Flip is probably one of the maneuvers that our approach can do very well, but human pilots find very challenging. It basically entails doing a high speed power loop by always looking backward. It is super challenging for humans, since they don’t see where they’re going and have problems in estimating their speed. For our approach the maneuver is no problem at all, since we can estimate forward velocities as well as backward velocities.

What are the limits to the performance of this system?

At the moment the main limitation is the maneuver duration. We never trained a controller that could perform maneuvers longer than 20 seconds. In the future, we plan to address this limitation and train general controllers which can fly in that agile way for significantly longer with relatively small drift. In this way, we could start being competitive against human pilots in drone racing competitions.

Can you talk about how the techniques developed here could be applied beyond drone acrobatics?

The current approach allows us to do acrobatics and agile flight in free space. We are now working to perform agile flight in cluttered environments, which requires a higher degree of understanding of the surrounding with respect to this project. Drone acrobatics is of course only an example application. We selected it because it makes a stress test of the controller performance. However, several other applications which require fast and agile flight can benefit from our approach. Examples are delivery (we want our Amazon packets always faster, don’t we?), search and rescue, or inspection. Going faster allows us to cover more space in less time, saving battery costs. Indeed, agile flight has very similar battery consumption of slow hovering for an autonomous drone.

“Deep Drone Acrobatics,” by Elia Kaufmann, Antonio Loquercio, René Ranftl, Matthias Müller, Vladlen Koltun, and Davide Scaramuzza from the Robotics and Perception Group at the University of Zurich and ETH Zurich, and Intel’s Intelligent Systems Lab, was presented at RSS 2020. Continue reading

Posted in Human Robots

#437614 Video Friday: Poimo Is a Portable ...

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 25-29, 2020 – [Online]
ROS World 2020 – November 12, 2020 – [Online]
CYBATHLON 2020 – November 13-14, 2020 – [Online]
ICSR 2020 – November 14-16, 2020 – Golden, Colo., USA
Let us know if you have suggestions for next week, and enjoy today's videos.

Engineers at the University of California San Diego have built a squid-like robot that can swim untethered, propelling itself by generating jets of water. The robot carries its own power source inside its body. It can also carry a sensor, such as a camera, for underwater exploration.

[ UCSD ]

Thanks Ioana!

Shark Robotics, French and European leader in Unmanned Ground Vehicles, is announcing today a disinfection add-on for Boston Dynamics Spot robot, designed to fight the COVID-19 pandemic. The Spot robot with Shark’s purpose-built disinfection payload can decontaminate up to 2,000 m2 in 15 minutes, in any space that needs to be sanitized – such as hospitals, metro stations, offices, warehouses or facilities.

[ Shark Robotics ]

Here’s an update on the Poimo portable inflatable mobility project we wrote about a little while ago; while not strictly robotics, it seems like it holds some promise for rapidly developing different soft structures that robotics might find useful.

[ University of Tokyo ]

Thanks Ryuma!

Pretty cool that you can do useful force feedback teleop while video chatting through a “regular broadband Internet connection.” Although, what “regular” means to you is a bit subjective, right?

[ HEBI Robotics ]

Thanks Dave!

While NASA's Mars rover Perseverance travels through space toward the Red Planet, its nearly identical rover twin is hard at work on Earth. The vehicle system test bed (VSTB) rover named OPTIMISM is a full-scale engineering version of the Mars-bound rover. It is used to test hardware and software before the commands are sent up to the Perseverance rover.

[ NASA ]

Jacquard takes ordinary, familiar objects and enhances them with new digital abilities and experiences, while remaining true to their original purpose — like being your favorite jacket, backpack or a pair of shoes that you love to wear.

Our ambition is simple: to make life easier. By staying connected to your digital world, your things can do so much more. Skip a song by brushing your sleeve. Take a picture by tapping on a shoulder strap. Get reminded about the phone you left behind with a blink of light or a haptic buzz on your cuff.

[ Google ATAP ]

Should you attend the IROS 2020 workshop on “Planetary Exploration Robots: Challenges and Opportunities”? Of course you should!

[ Workshop ]

Kuka makes a lot of these videos where I can’t help but think that if they put as much effort into programming the robot as they did into producing the video, the result would be much more impressive.

[ Kuka ]

The Colorado School of Mines is one of the first customers to buy a Spot robot from Boston Dynamics to help with robotics research. Watch as scientists take Spot into the school's mine for the first time.

[ HCR ] via [ CNET ]

A very interesting soft(ish) actuator from Ayato Kanada at Kyushu University's Control Engineering Lab.

A flexible ultrasonic motor (FUSM), which generates linear motion as a novel soft actuator. This motor consists of a single metal cube stator with a hole and an elastic elongated coil spring inserted into the hole. When voltages are applied to piezoelectric plates on the stator, the coil spring moves back and forward as a linear slider. In the FUSM that uses the friction drive as the principle, the most important parameter for optimizing its output is the preload between the stator and slider. The coil spring has a slightly larger diameter than the stator hole and generates the preload by expanding in a radial direction. The coil springs act not only as a flexible slider but also as a resistive positional sensor. Changes in the resistance between the stator and the coil spring end are converted to a voltage and used for position detection.

[ Control Engineering Lab ]

Thanks Ayato!

We show how to use the limbs of a quadruped robot to identify fine-grained soil, representative for Martian regolith.

[ Paper ] via [ ANYmal Research ]

PR2 is serving breakfast and cleaning up afterwards. It’s slow, but all you have to do is eat and leave.

That poor PR2 is a little more naked than it's probably comfortable with.

[ EASE ]

NVIDIA researchers present a hierarchical framework that combines model-based control and reinforcement learning (RL) to synthesize robust controllers for a quadruped robot (the Unitree Laikago).

[ NVIDIA ]

What's interesting about this assembly task is that the robot is using its arm only for positioning, and doing the actual assembly with just fingers.

[ RC2L ]

In this electronics assembly application, Kawasaki's cobot duAro2 uses a tool changing station to tackle a multitude of tasks and assemble different CPU models.

Okay but can it apply thermal paste to a CPU in the right way? Personally, I find that impossible.

[ Kawasaki ]

You only need to watch this video long enough to appreciate the concept of putting a robot on a robot.

[ Impress ]

In this lecture, we’ll hear from the man behind one of the biggest robotics companies in the world, Boston Dynamics, whose robotic dog, Spot, has been used to encourage social distancing in Singapore and is now getting ready for FDA approval to be able to measure patients’ vital signs in hospitals.

[ Alan Turing Institute ]

Greg Kahn from UC Berkeley wrote in to share his recent dissertation talk on “Mobile Robot Learning.”

In order to create mobile robots that can autonomously navigate real-world environments, we need generalizable perception and control systems that can reason about the outcomes of navigational decisions. Learning-based methods, in which the robot learns to navigate by observing the outcomes of navigational decisions in the real world, offer considerable promise for obtaining these intelligent navigation systems. However, there are many challenges impeding mobile robots from autonomously learning to act in the real-world, in particular (1) sample-efficiency–how to learn using a limited amount of data? (2) supervision–how to tell the robot what to do? and (3) safety–how to ensure the robot and environment are not damaged or destroyed during learning? In this talk, I will present deep reinforcement learning methods for addressing these real world mobile robot learning challenges and show results which enable ground and aerial robots to navigate in complex indoor and outdoor environments.

[ UC Berkeley ]

Thanks Greg!

Leila Takayama from UC Santa Cruz (and previously Google X and Willow Garage) gives a talk entitled “Toward a more human-centered future of robotics.”

Robots are no longer only in outer space, in factory cages, or in our imaginations. We interact with robotic agents when withdrawing cash from bank ATMs, driving cars with adaptive cruise control, and tuning our smart home thermostats. In the moment of those interactions with robotic agents, we behave in ways that do not necessarily align with the rational belief that robots are just plain machines. Through a combination of controlled experiments and field studies, we use theories and concepts from the social sciences to explore ways that human and robotic agents come together, including how people interact with personal robots and how people interact through telepresence robots. Together, we will explore topics and raise questions about the psychology of human-robot interaction and how we could invent a future of a more human-centered robotics that we actually want to live in.

[ Leila Takayama ]

Roboticist and stand-up comedian Naomi Fitter from Oregon State University gives a talk on “Everything I Know about Telepresence.”

Telepresence robots hold promise to connect people by providing videoconferencing and navigation abilities in far-away environments. At the same time, the impacts of current commercial telepresence robots are not well understood, and circumstances of robot use including internet connection stability, odd personalizations, and interpersonal relationship between a robot operator and people co-located with the robot can overshadow the benefit of the robot itself. And although the idea of telepresence robots has been around for over two decades, available nonverbal expressive abilities through telepresence robots are limited, and suitable operator user interfaces for the robot (for example, controls that allow for the operator to hold a conversation and move the robot simultaneously) remain elusive. So where should we be using telepresence robots? Are there any pitfalls to watch out for? What do we know about potential robot expressivity and user interfaces? This talk will cover my attempts to address these questions and ways in which the robotics research community can build off of this work

[ Talking Robotics ] Continue reading

Posted in Human Robots

#437592 Coordinated Robotics Wins DARPA SubT ...

DARPA held the Virtual Cave Circuit event of the Subterranean Challenge on Tuesday in the form of a several hour-long livestream. We got to watch (along with all of the competing teams) as virtual robots explored virtual caves fully autonomously, dodging rockfalls, spotting artifacts, scoring points, and sometimes running into stuff and falling over.

Expert commentary was provided by DARPA, and we were able to watch multiple teams running at once, skipping from highlight to highlight. It was really very well done (you can watch an archive of the entire stream here), but they made us wait until the very end to learn who won: First place went to Coordinated Robotics, with BARCS taking second, and third place going to newcomer Team Dynamo.

Huge congratulations to Coordinated Robotics! It’s worth pointing out that the top three teams were separated by an incredibly small handful of points, and on a slightly different day, with slightly different artifact positions, any of them could have come out on top. This doesn’t diminish Coordinated Robotics’ victory in the least—it means that the competition was fierce, and that the problem of autonomous cave exploration with robots has been solved (virtually, at least) in several different but effective ways.

We know Coordinated Robotics pretty well at this point, but here’s an introduction video:

You heard that right—Coordinated Robotics is just Kevin Knoedler, all by himself. This would be astonishing, if we weren’t already familiar with Kevin’s abilities: He won NASA’s virtual Space Robotics Challenge by himself in 2017, and Coordinated Robotics placed first in the DARPA SubT Virtual Tunnel Circuit and second in the Virtual Urban Circuit. We asked Kevin how he managed to do so spectacularly well (again), and here’s what he told us:

IEEE Spectrum: Can you describe what it was like to watch your team of robots on the live stream, and to see them score the most points?

Kevin Knoedler: It was exciting and stressful watching the live stream. It was exciting as the top few scores were quite close for the cave circuit. It was stressful because I started out behind and worked my way up, but did not do well on the final world. Luckily, not doing well on the first and last worlds was offset by better scores on many of the runs in between. DARPA did a very nice job with their live stream of the cave circuit results.

How did you decide on the makeup of your team, and on what sensors to use?

To decide on the makeup of the team I experimented with quite a few different vehicles. I had a lot of trouble with the X2 and other small ground vehicles flipping over. Based on that I looked at the larger ground vehicles that also had a sensor capable of identifying drop-offs. The vehicles that met those criteria for me were the Marble HD2, Marble Husky, Ozbot ATR, and the Absolem. Of those ground vehicles I went with the Marble HD2. It had a downward looking depth camera that I could use to detect drop-offs and was much more stable on the varied terrain than the X2. I had used the X3 aerial vehicle before and so that was my first choice for an aerial platform.

What were some things that you learned in Tunnel and Urban that you were able to incorporate into your strategy for Cave?

In the Tunnel circuit I had learned a strategy to use ground vehicles and in the Urban circuit I had learned a strategy to use aerial vehicles. At a high level that was the biggest thing I learned from the previous circuits that I was able to apply to the Cave circuit. At a lower level I was able to apply many of the development and testing strategies from the previous circuits to the Cave circuit.

What aspect of the cave environment was most challenging for your robots?

I would say it wasn't just one aspect of the cave environment that was challenging for the robots. There were quite a few challenging aspects of the cave environment. For the ground vehicles there were frequently paths that looked good as the robot started on the path, but turned into drop-offs or difficult boulder crawls. While it was fun to see the robot plan well enough to slowly execute paths over the boulders, I was wishing that the robot was smart enough to try a different path rather than wasting so much time crawling over the large boulders. For the aerial vehicles the combination of tight paths along with large vertical spaces was the biggest challenge in the environment. The large open vertical areas were particularly challenging for my aerial robots. They could easily lose track of their position without enough nearby features to track and it was challenging to find the correct path in and out of such large vertical areas.

How will you be preparing for the SubT Final?

To prepare for the SubT Final the vehicles will be getting a lot smarter. The ground vehicles will be better at navigation and communicating with one another. The aerial vehicles will be better able to handle large vertical areas both from a positioning and a planning point of view. Finally, all of the vehicles will do a better job coordinating what areas have been explored and what areas have good leads for further exploration.

Image: DARPA

The final score for the DARPA SubT Cave Circuit virtual competition.

We also had a chance to ask SubT program manager Tim Chung a few questions at yesterday’s post-event press conference, about the course itself and what he thinks teams should have learned from the competition:

IEEE Spectrum: Having looked through some real caves, can you give some examples of some of the most significant differences between this simulation and real caves? And with the enormous variety of caves out there, how generalizable are the solutions that teams came up with?

Tim Chung: Many of the caves that I’ve had to crawl through and gotten bumps and scrapes from had a couple of different features that I’ll highlight. The first is the variations in moisture— a lot of these caves were naturally formed with streams and such, so many of the caves we went to had significant mud, flowing water, and such. And so one of the things we're not capturing in the SubT simulator is explicitly anything that would submerge the robots, or otherwise short any of their systems. So from that perspective, that's one difference that's certainly notable.

And then the other difference I think is the granularity of the terrain, whether it's rubble, sand, or just raw dirt, friction coefficients are all across the board, and I think that's one of the things that any terrestrial simulator will both struggle with and potentially benefit from— that is, terramechanics simulation abilities. Given the emphasis on mobility in the SubT simulation, we’re capturing just a sliver of the complexity of terramechanics, but I think that's probably another take away that you'll certainly see— where there’s that distinction between physical and virtual technologies.

To answer your second question about generalizability— that’s the multi-million dollar question! It’s definitely at the crux of why we have eight diverse worlds, both in size verticality, dimensions, constraint passageways, etc. But this is eight out of countless variations, and the goal of course is to be able to investigate what those key dependencies are. What I'll say is that the out of the seventy three different virtual cave tiles, which are the building blocks that make up these virtual worlds, quite a number of them were not only inspired by real world caves, but were specifically designed so that we can essentially use these tiles as unit tests going forward. So, if I want to simulate vertical inclines, here are the tiles that are the vertical vertical unit tests for robots, and that’s how we’re trying to to think through how to tease out that generalizability factor.

What are some observations from this event that you think systems track teams should pay attention to as they prepare for the final event?

One of the key things about the virtual competition is that you submit your software, and that's it. So you have to design everything from state management to failure mode triage, really thinking about what could go wrong and then building out your autonomous capabilities either to react to some of those conditions, or to anticipate them. And to be honest I think that the humans in the loop that we have in the systems competition really are key enablers of their capability, but also could someday (if not already) be a crutch that we might not be able to develop.

Thinking through some of the failure modes in a fully autonomous software deployed setting are going to be incredibly valuable for the systems competitors, so that for example the human supervisor doesn't have to worry about those failure modes as much, or can respond in a more supervisory way rather than trying to joystick the robot around. I think that's going to be one of the greatest impacts, thinking through what it means to send these robots off to autonomously get you the information you need and complete the mission

This isn’t to say that the humans aren't going to be useful and continue to play a role of course, but I think this shifting of the role of the human supervisor from being a state manager to being more of a tactical commander will dramatically highlight the impact of the virtual side on the systems side.

What, if anything, should we take away from one person teams being able to do so consistently well in the virtual circuit?

It’s a really interesting question. I think part of it has to do with systems integration versus software integration. There's something to be said for the richness of the technologies that can be developed, and how many people it requires to be able to develop some of those technologies. With the systems competitors, having one person try to build, manage, deploy, service, and operate all of those robots is still functionally quite challenging, whereas in the virtual competition, it really is a software deployment more than anything else. And so I think the commonality of single person teams may just be a virtue of the virtual competition not having some of those person-intensive requirements.

In terms of their strong performance, I give credit to all of these really talented folks who are taking upon themselves to jump into the competitor pool and see how well they do, and I think that just goes to show you that whether you're one person or ten people people or a hundred people on a team, a good idea translated and executed well really goes a long way.

Looking ahead, teams have a year to prepare for the final event, which is still scheduled to be held sometime in fall 2021. And even though there was no cave event for systems track teams, the fact that the final event will be a combination of tunnel, urban, and cave circuits means that systems track teams have been figuring out how to get their robots to work in caves anyway, and we’ll be bringing you some of their stories over the next few weeks.

[ DARPA SubT ] Continue reading

Posted in Human Robots

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

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

#437466 How Future AI Could Recognize a Kangaroo ...

AI is continuously taking on new challenges, from detecting deepfakes (which, incidentally, are also made using AI) to winning at poker to giving synthetic biology experiments a boost. These impressive feats result partly from the huge datasets the systems are trained on. That training is costly and time-consuming, and it yields AIs that can really only do one thing well.

For example, to train an AI to differentiate between a picture of a dog and one of a cat, it’s fed thousands—if not millions—of labeled images of dogs and cats. A child, on the other hand, can see a dog or cat just once or twice and remember which is which. How can we make AIs learn more like children do?

A team at the University of Waterloo in Ontario has an answer: change the way AIs are trained.

Here’s the thing about the datasets normally used to train AI—besides being huge, they’re highly specific. A picture of a dog can only be a picture of a dog, right? But what about a really small dog with a long-ish tail? That sort of dog, while still being a dog, looks more like a cat than, say, a fully-grown Golden Retriever.

It’s this concept that the Waterloo team’s methodology is based on. They described their work in a paper published on the pre-print (or non-peer-reviewed) server arXiv last month. Teaching an AI system to identify a new class of objects using just one example is what they call “one-shot learning.” But they take it a step further, focusing on “less than one shot learning,” or LO-shot learning for short.

LO-shot learning consists of a system learning to classify various categories based on a number of examples that’s smaller than the number of categories. That’s not the most straightforward concept to wrap your head around, so let’s go back to the dogs and cats example. Say you want to teach an AI to identify dogs, cats, and kangaroos. How could that possibly be done without several clear examples of each animal?

The key, the Waterloo team says, is in what they call soft labels. Unlike hard labels, which label a data point as belonging to one specific class, soft labels tease out the relationship or degree of similarity between that data point and multiple classes. In the case of an AI trained on only dogs and cats, a third class of objects, say, kangaroos, might be described as 60 percent like a dog and 40 percent like a cat (I know—kangaroos probably aren’t the best animal to have thrown in as a third category).

“Soft labels can be used to represent training sets using fewer prototypes than there are classes, achieving large increases in sample efficiency over regular (hard-label) prototypes,” the paper says. Translation? Tell an AI a kangaroo is some fraction cat and some fraction dog—both of which it’s seen and knows well—and it’ll be able to identify a kangaroo without ever having seen one.

If the soft labels are nuanced enough, you could theoretically teach an AI to identify a large number of categories based on a much smaller number of training examples.

The paper’s authors use a simple machine learning algorithm called k-nearest neighbors (kNN) to explore this idea more in depth. The algorithm operates under the assumption that similar things are most likely to exist near each other; if you go to a dog park, there will be lots of dogs but no cats or kangaroos. Go to the Australian grasslands and there’ll be kangaroos but no cats or dogs. And so on.

To train a kNN algorithm to differentiate between categories, you choose specific features to represent each category (i.e. for animals you could use weight or size as a feature). With one feature on the x-axis and the other on the y-axis, the algorithm creates a graph where data points that are similar to each other are clustered near each other. A line down the center divides the categories, and it’s pretty straightforward for the algorithm to discern which side of the line new data points should fall on.

The Waterloo team kept it simple and used plots of color on a 2D graph. Using the colors and their locations on the graphs, the team created synthetic data sets and accompanying soft labels. One of the more simplistic graphs is pictured below, along with soft labels in the form of pie charts.

Image Credit: Ilia Sucholutsky & Matthias Schonlau
When the team had the algorithm plot the boundary lines of the different colors based on these soft labels, it was able to split the plot up into more colors than the number of data points it was given in the soft labels.

While the results are encouraging, the team acknowledges that they’re just the first step, and there’s much more exploration of this concept yet to be done. The kNN algorithm is one of the least complex models out there; what might happen when LO-shot learning is applied to a far more complex algorithm? Also, to apply it, you still need to distill a larger dataset down into soft labels.

One idea the team is already working on is having other algorithms generate the soft labels for the algorithm that’s going to be trained using LO-shot; manually designing soft labels won’t always be as easy as splitting up some pie charts into different colors.

LO-shot’s potential for reducing the amount of training data needed to yield working AI systems is promising. Besides reducing the cost and the time required to train new models, the method could also make AI more accessible to industries, companies, or individuals who don’t have access to large datasets—an important step for democratization of AI.

Image Credit: pen_ash from Pixabay Continue reading

Posted in Human Robots