Tag Archives: d

#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

#437620 The Trillion-Transistor Chip That Just ...

The history of computer chips is a thrilling tale of extreme miniaturization.

The smaller, the better is a trend that’s given birth to the digital world as we know it. So, why on earth would you want to reverse course and make chips a lot bigger? Well, while there’s no particularly good reason to have a chip the size of an iPad in an iPad, such a chip may prove to be genius for more specific uses, like artificial intelligence or simulations of the physical world.

At least, that’s what Cerebras, the maker of the biggest computer chip in the world, is hoping.

The Cerebras Wafer-Scale Engine is massive any way you slice it. The chip is 8.5 inches to a side and houses 1.2 trillion transistors. The next biggest chip, NVIDIA’s A100 GPU, measures an inch to a side and has a mere 54 billion transistors. The former is new, largely untested and, so far, one-of-a-kind. The latter is well-loved, mass-produced, and has taken over the world of AI and supercomputing in the last decade.

So can Goliath flip the script on David? Cerebras is on a mission to find out.

Big Chips Beyond AI
When Cerebras first came out of stealth last year, the company said it could significantly speed up the training of deep learning models.

Since then, the WSE has made its way into a handful of supercomputing labs, where the company’s customers are putting it through its paces. One of those labs, the National Energy Technology Laboratory, is looking to see what it can do beyond AI.

So, in a recent trial, researchers pitted the chip—which is housed in an all-in-one system about the size of a dorm room mini-fridge called the CS-1—against a supercomputer in a fluid dynamics simulation. Simulating the movement of fluids is a common supercomputer application useful for solving complex problems like weather forecasting and airplane wing design.

The trial was described in a preprint paper written by a team led by Cerebras’s Michael James and NETL’s Dirk Van Essendelft and presented at the supercomputing conference SC20 this week. The team said the CS-1 completed a simulation of combustion in a power plant roughly 200 times faster than it took the Joule 2.0 supercomputer to do a similar task.

The CS-1 was actually faster-than-real-time. As Cerebrus wrote in a blog post, “It can tell you what is going to happen in the future faster than the laws of physics produce the same result.”

The researchers said the CS-1’s performance couldn’t be matched by any number of CPUs and GPUs. And CEO and cofounder Andrew Feldman told VentureBeat that would be true “no matter how large the supercomputer is.” At a point, scaling a supercomputer like Joule no longer produces better results in this kind of problem. That’s why Joule’s simulation speed peaked at 16,384 cores, a fraction of its total 86,400 cores.

A comparison of the two machines drives the point home. Joule is the 81st fastest supercomputer in the world, takes up dozens of server racks, consumes up to 450 kilowatts of power, and required tens of millions of dollars to build. The CS-1, by comparison, fits in a third of a server rack, consumes 20 kilowatts of power, and sells for a few million dollars.

While the task is niche (but useful) and the problem well-suited to the CS-1, it’s still a pretty stunning result. So how’d they pull it off? It’s all in the design.

Cut the Commute
Computer chips begin life on a big piece of silicon called a wafer. Multiple chips are etched onto the same wafer and then the wafer is cut into individual chips. While the WSE is also etched onto a silicon wafer, the wafer is left intact as a single, operating unit. This wafer-scale chip contains almost 400,000 processing cores. Each core is connected to its own dedicated memory and its four neighboring cores.

Putting that many cores on a single chip and giving them their own memory is why the WSE is bigger; it’s also why, in this case, it’s better.

Most large-scale computing tasks depend on massively parallel processing. Researchers distribute the task among hundreds or thousands of chips. The chips need to work in concert, so they’re in constant communication, shuttling information back and forth. A similar process takes place within each chip, as information moves between processor cores, which are doing the calculations, and shared memory to store the results.

It’s a little like an old-timey company that does all its business on paper.

The company uses couriers to send and collect documents from other branches and archives across town. The couriers know the best routes through the city, but the trips take some minimum amount of time determined by the distance between the branches and archives, the courier’s top speed, and how many other couriers are on the road. In short, distance and traffic slow things down.

Now, imagine the company builds a brand new gleaming skyscraper. Every branch is moved into the new building and every worker gets a small filing cabinet in their office to store documents. Now any document they need can be stored and retrieved in the time it takes to step across the office or down the hall to their neighbor’s office. The information commute has all but disappeared. Everything’s in the same house.

Cerebras’s megachip is a bit like that skyscraper. The way it shuttles information—aided further by its specially tailored compiling software—is far more efficient compared to a traditional supercomputer that needs to network a ton of traditional chips.

Simulating the World as It Unfolds
It’s worth noting the chip can only handle problems small enough to fit on the wafer. But such problems may have quite practical applications because of the machine’s ability to do high-fidelity simulation in real-time. The authors note, for example, the machine should in theory be able to accurately simulate the air flow around a helicopter trying to land on a flight deck and semi-automate the process—something not possible with traditional chips.

Another opportunity, they note, would be to use a simulation as input to train a neural network also residing on the chip. In an intriguing and related example, a Caltech machine learning technique recently proved to be 1,000 times faster at solving the same kind of partial differential equations at play here to simulate fluid dynamics.

They also note that improvements in the chip (and others like it, should they arrive) will push back the limits of what can be accomplished. Already, Cerebras has teased the release of its next-generation chip, which will have 2.6 trillion transistors, 850,00 cores, and more than double the memory.

Of course, it still remains to be seen whether wafer-scale computing really takes off. The idea has been around for decades, but Cerebras is the first to pursue it seriously. Clearly, they believe they’ve solved the problem in a way that’s useful and economical.

Other new architectures are also being pursued in the lab. Memristor-based neuromorphic chips, for example, mimic the brain by putting processing and memory into individual transistor-like components. And of course, quantum computers are in a separate lane, but tackle similar problems.

It could be that one of these technologies eventually rises to rule them all. Or, and this seems just as likely, computing may splinter into a bizarre quilt of radical chips, all stitched together to make the most of each depending on the situation.

Image credit: Cerebras Continue reading

Posted in Human Robots

#437610 How Intel’s OpenBot Wants to Make ...

You could make a pretty persuasive argument that the smartphone represents the single fastest area of technological progress we’re going to experience for the foreseeable future. Every six months or so, there’s something with better sensors, more computing power, and faster connectivity. Many different areas of robotics are benefiting from this on a component level, but over at Intel Labs, they’re taking a more direct approach with a project called OpenBot that turns US $50 worth of hardware and your phone into a mobile robot that can support “advanced robotics workloads such as person following and real-time autonomous navigation in unstructured environments.”

This work aims to address two key challenges in robotics: accessibility and scalability. Smartphones are ubiquitous and are becoming more powerful by the year. We have developed a combination of hardware and software that turns smartphones into robots. The resulting robots are inexpensive but capable. Our experiments have shown that a $50 robot body powered by a smartphone is capable of person following and real-time autonomous navigation. We hope that the presented work will open new opportunities for education and large-scale learning via thousands of low-cost robots deployed around the world.

Smartphones point to many possibilities for robotics that we have not yet exploited. For example, smartphones also provide a microphone, speaker, and screen, which are not commonly found on existing navigation robots. These may enable research and applications at the confluence of human-robot interaction and natural language processing. We also expect the basic ideas presented in this work to extend to other forms of robot embodiment, such as manipulators, aerial vehicles, and watercraft.

One of the interesting things about this idea is how not-new it is. The highest profile phone robot was likely the $150 Romo, from Romotive, which raised a not-insignificant amount of money on Kickstarter in 2012 and 2013 for a little mobile chassis that accepted one of three different iPhone models and could be controlled via another device or operated somewhat autonomously. It featured “computer vision, autonomous navigation, and facial recognition” capabilities, but was really designed to be a toy. Lack of compatibility hampered Romo a bit, and there wasn’t a lot that it could actually do once the novelty wore off.

As impressive as smartphone hardware was in a robotics context (even back in 2013), we’re obviously way, way beyond that now, and OpenBot figures that smartphones now have enough clout and connectivity that turning them into mobile robots is a good idea. You know, again. We asked Intel Labs’ Matthias Muller why now was the right time to launch OpenBot, and he mentioned things like the existence of a large maker community with broad access to 3D printing as well as open source software that makes broader development easier.

And of course, there’s the smartphone hardware: “Smartphones have become extremely powerful and feature dedicated AI processors in addition to CPUs and GPUs,” says Mueller. “Almost everyone owns a very capable smartphone now. There has been a big boost in sensor performance, especially in cameras, and a lot of the recent developments for VR applications are well aligned with robotic requirements for state estimation.” OpenBot has been tested with 10 recent Android phones, and since camera placement tends to be similar and USB-C is becoming the charging and communications standard, compatibility is less of an issue nowadays.

Image: OpenBot

Intel researchers created this table comparing OpenBot to other wheeled robot platforms, including Amazon’s DeepRacer, MIT’s Duckiebot, iRobot’s Create-2, and Thymio. The top group includes robots based on RC trucks; the bottom group includes navigation robots for deployment at scale and in education. Note that the cost of the smartphone needed for OpenBot is not included in this comparison.

If you’d like an OpenBot of your own, you don’t need to know all that much about robotics hardware or software. For the hardware, you probably need some basic mechanical and electronics experience—think Arduino project level. The software is a little more complicated; there’s a pretty good walkthrough to get some relatively sophisticated behaviors (like autonomous person following) up and running, but things rapidly degenerate into a command line interface that could be intimidating for new users. We did ask about why OpenBot isn’t ROS-based to leverage the robustness and reach of that community, and Muller said that ROS “adds unnecessary overhead,” although “if someone insists on using ROS with OpenBot, it should not be very difficult.”

Without building OpenBot to explicitly be part of an existing ecosystem, the challenge going forward is to make sure that the project is consistently supported, lest it wither and die like so many similar robotics projects have before it. “We are committed to the OpenBot project and will do our best to maintain it,” Mueller assures us. “We have a good track record. Other projects from our group (e.g. CARLA, Open3D, etc.) have also been maintained for several years now.” The inherently open source nature of the project certainly helps, although it can be tricky to rely too much on community contributions, especially when something like this is first starting out.

The OpenBot folks at Intel, we’re told, are already working on a “bigger, faster and more powerful robot body that will be suitable for mass production,” which would certainly help entice more people into giving this thing a go. They’ll also be focusing on documentation, which is probably the most important but least exciting part about building a low-cost community focused platform like this. And as soon as they’ve put together a way for us actual novices to turn our phones into robots that can do cool stuff for cheap, we’ll definitely let you know. Continue reading

Posted in Human Robots

#437585 Dart-Shooting Drone Attacks Trees for ...

We all know how robots are great at going to places where you can’t (or shouldn’t) send a human. We also know how robots are great at doing repetitive tasks. These characteristics have the potential to make robots ideal for setting up wireless sensor networks in hazardous environments—that is, they could deploy a whole bunch of self-contained sensor nodes that create a network that can monitor a very large area for a very long time.

When it comes to using drones to set up sensor networks, you’ve generally got two options: A drone that just drops sensors on the ground (easy but inaccurate and limited locations), or using a drone with some sort of manipulator on it to stick sensors in specific places (complicated and risky). A third option, under development by researchers at Imperial College London’s Aerial Robotics Lab, provides the accuracy of direct contact with the safety and ease of use of passive dropping by instead using the drone as a launching platform for laser-aimed, sensor-equipped darts.

These darts (which the researchers refer to as aerodynamically stabilized, spine-equipped sensor pods) can embed themselves in relatively soft targets from up to 4 meters away with an accuracy of about 10 centimeters after being fired from a spring-loaded launcher. They’re not quite as accurate as a drone with a manipulator, but it’s pretty good, and the drone can maintain a safe distance from the surface that it’s trying to add a sensor to. Obviously, the spine is only going to work on things like wood, but the researchers point out that there are plenty of attachment mechanisms that could be used, including magnets, adhesives, chemical bonding, or microspines.

Indoor tests using magnets showed the system to be quite reliable, but at close range (within a meter of the target) the darts sometimes bounced off rather than sticking. From between 1 and 4 meters away, the darts stuck between 90 and 100 percent of the time. Initial outdoor tests were also successful, although the system was under manual control. The researchers say that “regular and safe operations should be carried out autonomously,” which, yeah, you’d just have to deal with all of the extra sensing and hardware required to autonomously fly beneath the canopy of a forest. That’s happening next, as the researchers plan to add “vision state estimation and positioning, as well as a depth sensor” to avoid some trees and fire sensors into others.

And if all of that goes well, they’ll consider trying to get each drone to carry multiple darts. Look out, trees: You’re about to be pierced for science.

“Unmanned Aerial Sensor Placement for Cluttered Environments,” by André Farinha, Raphael Zufferey, Peter Zheng, Sophie F. Armanini, and Mirko Kovac from Imperial College London, was published in IEEE Robotics and Automation Letters.

< Back to IEEE Journal Watch Continue reading

Posted in Human Robots

#437579 Disney Research Makes Robotic Gaze ...

While it’s not totally clear to what extent human-like robots are better than conventional robots for most applications, one area I’m personally comfortable with them is entertainment. The folks over at Disney Research, who are all about entertainment, have been working on this sort of thing for a very long time, and some of their animatronic attractions are actually quite impressive.

The next step for Disney is to make its animatronic figures, which currently feature scripted behaviors, to perform in an interactive manner with visitors. The challenge is that this is where you start to get into potential Uncanny Valley territory, which is what happens when you try to create “the illusion of life,” which is what Disney (they explicitly say) is trying to do.

In a paper presented at IROS this month, a team from Disney Research, Caltech, University of Illinois at Urbana-Champaign, and Walt Disney Imagineering is trying to nail that illusion of life with a single, and perhaps most important, social cue: eye gaze.

Before you watch this video, keep in mind that you’re watching a specific character, as Disney describes:

The robot character plays an elderly man reading a book, perhaps in a library or on a park bench. He has difficulty hearing and his eyesight is in decline. Even so, he is constantly distracted from reading by people passing by or coming up to greet him. Most times, he glances at people moving quickly in the distance, but as people encroach into his personal space, he will stare with disapproval for the interruption, or provide those that are familiar to him with friendly acknowledgment.

What, exactly, does “lifelike” mean in the context of robotic gaze? The paper abstract describes the goal as “[seeking] to create an interaction which demonstrates the illusion of life.” I suppose you could think of it like a sort of old-fashioned Turing test focused on gaze: If the gaze of this robot cannot be distinguished from the gaze of a human, then victory, that’s lifelike. And critically, we’re talking about mutual gaze here—not just a robot gazing off into the distance, but you looking deep into the eyes of this robot and it looking right back at you just like a human would. Or, just like some humans would.

The approach that Disney is using is more animation-y than biology-y or psychology-y. In other words, they’re not trying to figure out what’s going on in our brains to make our eyes move the way that they do when we’re looking at other people and basing their control system on that, but instead, Disney just wants it to look right. This “visual appeal” approach is totally fine, and there’s been an enormous amount of human-robot interaction (HRI) research behind it already, albeit usually with less explicitly human-like platforms. And speaking of human-like platforms, the hardware is a “custom Walt Disney Imagineering Audio-Animatronics bust,” which has DoFs that include neck, eyes, eyelids, and eyebrows.

In order to decide on gaze motions, the system first identifies a person to target with its attention using an RGB-D camera. If more than one person is visible, the system calculates a curiosity score for each, currently simplified to be based on how much motion it sees. Depending on which person that the robot can see has the highest curiosity score, the system will choose from a variety of high level gaze behavior states, including:

Read: The Read state can be considered the “default” state of the character. When not executing another state, the robot character will return to the Read state. Here, the character will appear to read a book located at torso level.

Glance: A transition to the Glance state from the Read or Engage states occurs when the attention engine indicates that there is a stimuli with a curiosity score […] above a certain threshold.

Engage: The Engage state occurs when the attention engine indicates that there is a stimuli […] to meet a threshold and can be triggered from both Read and Glance states. This state causes the robot to gaze at the person-of-interest with both the eyes and head.

Acknowledge: The Acknowledge state is triggered from either Engage or Glance states when the person-of-interest is deemed to be familiar to the robot.

Running underneath these higher level behavior states are lower level motion behaviors like breathing, small head movements, eye blinking, and saccades (the quick eye movements that occur when people, or robots, look between two different focal points). The term for this hierarchical behavioral state layering is a subsumption architecture, which goes all the way back to Rodney Brooks’ work on robots like Genghis in the 1980s and Cog and Kismet in the ’90s, and it provides a way for more complex behaviors to emerge from a set of simple, decentralized low-level behaviors.

“25 years on Disney is using my subsumption architecture for humanoid eye control, better and smoother now than our 1995 implementations on Cog and Kismet.”
—Rodney Brooks, MIT emeritus professor

Brooks, an emeritus professor at MIT and, most recently, cofounder and CTO of Robust.ai, tweeted about the Disney project, saying: “People underestimate how long it takes to get from academic paper to real world robotics. 25 years on Disney is using my subsumption architecture for humanoid eye control, better and smoother now than our 1995 implementations on Cog and Kismet.”

From the paper:

Although originally intended for control of mobile robots, we find that the subsumption architecture, as presented in [17], lends itself as a framework for organizing animatronic behaviors. This is due to the analogous use of subsumption in human behavior: human psychomotor behavior can be intuitively modeled as layered behaviors with incoming sensory inputs, where higher behavioral levels are able to subsume lower behaviors. At the lowest level, we have involuntary movements such as heartbeats, breathing and blinking. However, higher behavioral responses can take over and control lower level behaviors, e.g., fight-or-flight response can induce faster heart rate and breathing. As our robot character is modeled after human morphology, mimicking biological behaviors through the use of a bottom-up approach is straightforward.

The result, as the video shows, appears to be quite good, although it’s hard to tell how it would all come together if the robot had more of, you know, a face. But it seems like you don’t necessarily need to have a lifelike humanoid robot to take advantage of this architecture in an HRI context—any robot that wants to make a gaze-based connection with a human could benefit from doing it in a more human-like way.

“Realistic and Interactive Robot Gaze,” by Matthew K.X.J. Pan, Sungjoon Choi, James Kennedy, Kyna McIntosh, Daniel Campos Zamora, Gunter Niemeyer, Joohyung Kim, Alexis Wieland, and David Christensen from Disney Research, California Institute of Technology, University of Illinois at Urbana-Champaign, and Walt Disney Imagineering, was presented at IROS 2020. You can find the full paper, along with a 13-minute video presentation, on the IROS on-demand conference website.

< Back to IEEE Journal Watch Continue reading

Posted in Human Robots