Tag Archives: master
#439053 Bipedal Robots Are Learning To Move With ...
Most humans are bipeds, but even the best of us are really only bipeds until things get tricky. While our legs may be our primary mobility system, there are lots of situations in which we leverage our arms as well, either passively to keep balance or actively when we put out a hand to steady ourselves on a nearby object. And despite how unstable bipedal robots tend to be, using anything besides legs for mobility has been a challenge in both software and hardware, a significant limitation in highly unstructured environments.
Roboticists from TUM in Germany (with support from the German Research Foundation) have recently given their humanoid robot LOLA some major upgrades to make this kind of multi-contact locomotion possible. While it’s still in the early stages, it’s already some of the most human-like bipedal locomotion we’ve seen.
It’s certainly possible for bipedal robots to walk over challenging terrain without using limbs for support, but I’m sure you can think of lots of times where using your arms to assist with your own bipedal mobility was a requirement. It’s not a requirement because your leg strength or coordination or sense of balance is bad, necessarily. It’s just that sometimes, you might find yourself walking across something that’s highly unstable or in a situation where the consequences of a stumble are exceptionally high. And it may not even matter how much sensing you do beforehand, and how careful you are with your footstep planning: there are limits to how much you can know about your environment beforehand, and that can result in having a really bad time of it. This is why using multi-contact locomotion, whether it’s planned in advance or not, is a useful skill for humans, and should be for robots, too.
As the video notes (and props for being explicit up front about it), this isn’t yet fully autonomous behavior, with foot positions and arm contact points set by hand in advance. But it’s not much of a stretch to see how everything could be done autonomously, since one of the really hard parts (using multiple contact points to dynamically balance a moving robot) is being done onboard and in real time.
Getting LOLA to be able to do this required a major overhaul in hardware as well as software. And Philipp Seiwald, who works with LOLA at TUM, was able to tell us more about it.
IEEE Spectrum: Can you summarize the changes to LOLA’s hardware that are required for multi-contact locomotion?
Philipp Seiwald: The original version of LOLA has been designed for fast biped walking. Although it had two arms, they were not meant to get into contact with the environment but rather to compensate for the dynamic effects of the feet during fast walking. Also, the torso had a relatively simple design that was fine for its original purpose; however, it was not conceived to withstand the high loads coming from the hands during multi-contact maneuvers. Thus, we redesigned the complete upper body of LOLA from scratch. Starting from the pelvis, the strength and stiffness of the torso have been increased. We used the finite element method to optimize critical parts to obtain maximum strength at minimum weight. Moreover, we added additional degrees of freedom to the arms to increase the hands' reachable workspace. The kinematic topology of the arms, i.e., the arrangement of joints and link lengths, has been obtained from an optimization that takes typical multi-contact scenarios into account.
Why is this an important problem for bipedal humanoid robots?
Maintaining balance during locomotion can be considered the primary goal of legged robots. Naturally, this task is more challenging for bipeds when compared to robots with four or even more legs. Although current high-end prototypes show impressive progress, humanoid robots still do not have the robustness and versatility they need for most real-world applications. With our research, we try to contribute to this field and help to push the limits further. Recently, we showed our latest work on walking over uneven terrain without multi-contact support. Although the robustness is already high, there still exist scenarios, such as walking on loose objects, where the robot's stabilization fails when using only foot contacts. The use of additional hand-environment support during this (comparatively) fast walking allows a further significant increase in robustness, i.e., the robot's capability to compensate disturbances, modeling errors, or inaccurate sensor input. Besides stabilization on uneven terrain, multi-contact locomotion also enables more complex motions, e.g., stepping over a tall obstacle or toe-only contacts, as shown in our latest multi-contact video.
How can LOLA decide whether a surface is suitable for multi-contact locomotion?
LOLA’s visual perception system is currently developed by our project partners from the Chair for Computer Aided Medical Procedures & Augmented Reality at the TUM. This system relies on a novel semantic Simultaneous Localization and Mapping (SLAM) pipeline that can robustly extract the scene's semantic components (like floor, walls, and objects therein) by merging multiple observations from different viewpoints and by inferring therefrom the underlying scene graph. This provides a reliable estimate of which scene parts can be used to support the locomotion, based on the assumption that certain structural elements such as walls are fixed, while chairs, for example, are not.
Also, the team plans to develop a specific dataset with annotations further describing the attributes of the object (such as roughness of the surface or its softness) and that will be used to master multi-contact locomotion in even more complex scenes. As of today, the vision and navigation system is not finished yet; thus, in our latest video, we used pre-defined footholds and contact points for the hands. However, within our collaboration, we are working towards a fully integrated and autonomous system.
Is LOLA capable of both proactive and reactive multi-contact locomotion?
The software framework of LOLA has a hierarchical structure. On the highest level, the vision system generates an environment model and estimates the 6D-pose of the robot in the scene. The walking pattern generator then uses this information to plan a dynamically feasible future motion that will lead LOLA to a target position defined by the user. On a lower level, the stabilization module modifies this plan to compensate for model errors or any kind of disturbance and keep overall balance. So our approach currently focuses on proactive multi-contact locomotion. However, we also plan to work on a more reactive behavior such that additional hand support can also be triggered by an unexpected disturbance instead of being planned in advance.
What are some examples of unique capabilities that you are working towards with LOLA?
One of the main goals for the research with LOLA remains fast, autonomous, and robust locomotion on complex, uneven terrain. We aim to reach a walking speed similar to humans. Currently, LOLA can do multi-contact locomotion and cross uneven terrain at a speed of 1.8 km/h, which is comparably fast for a biped robot but still slow for a human. On flat ground, LOLA's high-end hardware allows it to walk at a relatively high maximum speed of 3.38 km/h.
Fully autonomous multi-contact locomotion for a life-sized humanoid robot is a tough task. As algorithms get more complex, computation time increases, which often results in offline motion planning methods. For LOLA, we restrict ourselves to gaited multi-contact locomotion, which means that we try to preserve the core characteristics of bipedal gait and use the arms only for assistance. This allows us to use simplified models of the robot which lead to very efficient algorithms running in real-time and fully onboard.
A long-term scientific goal with LOLA is to understand essential components and control policies of human walking. LOLA's leg kinematics is relatively similar to the human body. Together with scientists from kinesiology, we try to identify similarities and differences between observed human walking and LOLA’s “engineered” walking gait. We hope this research leads, on the one hand, to new ideas for the control of bipeds, and on the other hand, shows via experiments on bipeds if biomechanical models for the human gait are correctly understood. For a comparison of control policies on uneven terrain, LOLA must be able to walk at comparable speeds, which also motivates our research on fast and robust walking.
While it makes sense why the researchers are using LOLA’s arms primarily to assist with a conventional biped gait, looking ahead a bit it’s interesting to think about how robots that we typically consider to be bipeds could potentially leverage their limbs for mobility in decidedly non-human ways.
We’re used to legged robots being one particular morphology, I guess because associating them with either humans or dogs or whatever is just a comfortable way to do it, but there’s no particular reason why a robot with four limbs has to choose between being a quadruped and being a biped with arms, or some hybrid between the two, depending on what its task is. The research being done with LOLA could be a step in that direction, and maybe a hand on the wall in that direction, too. Continue reading
Posted in Human Robots
#437905 New Deep Learning Method Helps Robots ...
One of the biggest things standing in the way of the robot revolution is their inability to adapt. That may be about to change though, thanks to a new approach that blends pre-learned skills on the fly to tackle new challenges.
Put a robot in a tightly-controlled environment and it can quickly surpass human performance at complex tasks, from building cars to playing table tennis. But throw these machines a curve ball and they’re in trouble—just check out this compilation of some of the world’s most advanced robots coming unstuck in the face of notoriously challenging obstacles like sand, steps, and doorways.
The reason robots tend to be so fragile is that the algorithms that control them are often manually designed. If they encounter a situation the designer didn’t think of, which is almost inevitable in the chaotic real world, then they simply don’t have the tools to react.
Rapid advances in AI have provided a potential workaround by letting robots learn how to carry out tasks instead of relying on hand-coded instructions. A particularly promising approach is deep reinforcement learning, where the robot interacts with its environment through a process of trial-and-error and is rewarded for carrying out the correct actions. Over many repetitions it can use this feedback to learn how to accomplish the task at hand.
But the approach requires huge amounts of data to solve even simple tasks. And most of the things we would want a robot to do are actually comprised of many smaller tasks—for instance, delivering a parcel involves learning how to pick an object up, how to walk, how to navigate, and how to pass an object to someone else, among other things.
Training all these sub-tasks simultaneously is hugely complex and far beyond the capabilities of most current AI systems, so many experiments so far have focused on narrow skills. Some have tried to train AI on multiple skills separately and then use an overarching system to flip between these expert sub-systems, but these approaches still can’t adapt to completely new challenges.
Building off this research, though, scientists have now created a new AI system that can blend together expert sub-systems specialized for a specific task. In a paper in Science Robotics, they explain how this allows a four-legged robot to improvise new skills and adapt to unfamiliar challenges in real time.
The technique, dubbed multi-expert learning architecture (MELA), relies on a two-stage training approach. First the researchers used a computer simulation to train two neural networks to carry out two separate tasks: trotting and recovering from a fall.
They then used the models these two networks learned as seeds for eight other neural networks specialized for more specific motor skills, like rolling over or turning left or right. The eight “expert networks” were trained simultaneously along with a “gating network,” which learns how to combine these experts to solve challenges.
Because the gating network synthesizes the expert networks rather than switching them on sequentially, MELA is able to come up with blends of different experts that allow it to tackle problems none could solve alone.
The authors liken the approach to training people in how to play soccer. You start out by getting them to do drills on individual skills like dribbling, passing, or shooting. Once they’ve mastered those, they can then intelligently combine them to deal with more dynamic situations in a real game.
After training the algorithm in simulation, the researchers uploaded it to a four-legged robot and subjected it to a battery of tests, both indoors and outdoors. The robot was able to adapt quickly to tricky surfaces like gravel or pebbles, and could quickly recover from being repeatedly pushed over before continuing on its way.
There’s still some way to go before the approach could be adapted for real-world commercially useful robots. For a start, MELA currently isn’t able to integrate visual perception or a sense of touch; it simply relies on feedback from the robot’s joints to tell it what’s going on around it. The more tasks you ask the robot to master, the more complex and time-consuming the training will get.
Nonetheless, the new approach points towards a promising way to make multi-skilled robots become more than the sum of their parts. As much fun as it is, it seems like laughing at compilations of clumsy robots may soon be a thing of the past.
Image Credit: Yang et al., Science Robotics Continue reading
Posted in Human Robots
#437872 AlphaFold Proves That AI Can Crack ...
Any successful implementation of artificial intelligence hinges on asking the right questions in the right way. That’s what the British AI company DeepMind (a subsidiary of Alphabet) accomplished when it used its neural network to tackle one of biology’s grand challenges, the protein-folding problem. Its neural net, known as AlphaFold, was able to predict the 3D structures of proteins based on their amino acid sequences with unprecedented accuracy.
AlphaFold’s predictions at the 14th Critical Assessment of protein Structure Prediction (CASP14) were accurate to within an atom’s width for most of the proteins. The competition consisted of blindly predicting the structure of proteins that have only recently been experimentally determined—with some still awaiting determination.
Called the building blocks of life, proteins consist of 20 different amino acids in various combinations and sequences. A protein's biological function is tied to its 3D structure. Therefore, knowledge of the final folded shape is essential to understanding how a specific protein works—such as how they interact with other biomolecules, how they may be controlled or modified, and so on. “Being able to predict structure from sequence is the first real step towards protein design,” says Janet M. Thornton, director emeritus of the European Bioinformatics Institute. It also has enormous benefits in understanding disease-causing pathogens. For instance, at the moment only about 18 of the 26 proteins in the SARS-CoV-2 virus are known.
Predicting a protein’s 3D structure is a computational nightmare. In 1969 Cyrus Levinthal estimated that there are 10300 possible conformational combinations for a single protein, which would take longer than the age of the known universe to evaluate by brute force calculation. AlphaFold can do it in a few days.
As scientific breakthroughs go, AlphaFold’s discovery is right up there with the likes of James Watson and Francis Crick’s DNA double-helix model, or, more recently, Jennifer Doudna and Emmanuelle Charpentier’s CRISPR-Cas9 genome editing technique.
How did a team that just a few years ago was teaching an AI to master a 3,000-year-old game end up training one to answer a question plaguing biologists for five decades? That, says Briana Brownell, data scientist and founder of the AI company PureStrategy, is the beauty of artificial intelligence: The same kind of algorithm can be used for very different things.
“Whenever you have a problem that you want to solve with AI,” she says, “you need to figure out how to get the right data into the model—and then the right sort of output that you can translate back into the real world.”
DeepMind’s success, she says, wasn’t so much a function of picking the right neural nets but rather “how they set up the problem in a sophisticated enough way that the neural network-based modeling [could] actually answer the question.”
AlphaFold showed promise in 2018, when DeepMind introduced a previous iteration of their AI at CASP13, achieving the highest accuracy among all participants. The team had trained its to model target shapes from scratch, without using previously solved proteins as templates.
For 2020 they deployed new deep learning architectures into the AI, using an attention-based model that was trained end-to-end. Attention in a deep learning network refers to a component that manages and quantifies the interdependence between the input and output elements, as well as between the input elements themselves.
The system was trained on public datasets of the approximately 170,000 known experimental protein structures in addition to databases with protein sequences of unknown structures.
“If you look at the difference between their entry two years ago and this one, the structure of the AI system was different,” says Brownell. “This time, they’ve figured out how to translate the real world into data … [and] created an output that could be translated back into the real world.”
Like any AI system, AlphaFold may need to contend with biases in the training data. For instance, Brownell says, AlphaFold is using available information about protein structure that has been measured in other ways. However, there are also many proteins with as yet unknown 3D structures. Therefore, she says, a bias could conceivably creep in toward those kinds of proteins that we have more structural data for.
Thornton says it’s difficult to predict how long it will take for AlphaFold’s breakthrough to translate into real-world applications.
“We only have experimental structures for about 10 per cent of the 20,000 proteins [in] the human body,” she says. “A powerful AI model could unveil the structures of the other 90 per cent.”
Apart from increasing our understanding of human biology and health, she adds, “it is the first real step toward… building proteins that fulfill a specific function. From protein therapeutics to biofuels or enzymes that eat plastic, the possibilities are endless.” Continue reading
Posted in Human Robots
#437809 Q&A: The Masterminds Behind ...
Illustration: iStockphoto
Getting a car to drive itself is undoubtedly the most ambitious commercial application of artificial intelligence (AI). The research project was kicked into life by the 2004 DARPA Urban Challenge and then taken up as a business proposition, first by Alphabet, and later by the big automakers.
The industry-wide effort vacuumed up many of the world’s best roboticists and set rival companies on a multibillion-dollar acquisitions spree. It also launched a cycle of hype that paraded ever more ambitious deadlines—the most famous of which, made by Alphabet’s Sergei Brin in 2012, was that full self-driving technology would be ready by 2017. Those deadlines have all been missed.
Much of the exhilaration was inspired by the seeming miracles that a new kind of AI—deep learning—was achieving in playing games, recognizing faces, and transliterating voices. Deep learning excels at tasks involving pattern recognition—a particular challenge for older, rule-based AI techniques. However, it now seems that deep learning will not soon master the other intellectual challenges of driving, such as anticipating what human beings might do.
Among the roboticists who have been involved from the start are Gill Pratt, the chief executive officer of Toyota Research Institute (TRI) , formerly a program manager at the Defense Advanced Research Projects Agency (DARPA); and Wolfram Burgard, vice president of automated driving technology for TRI and president of the IEEE Robotics and Automation Society. The duo spoke with IEEE Spectrum’s Philip Ross at TRI’s offices in Palo Alto, Calif.
This interview has been condensed and edited for clarity.
IEEE Spectrum: How does AI handle the various parts of the self-driving problem?
Photo: Toyota
Gill Pratt
Gill Pratt: There are three different systems that you need in a self-driving car: It starts with perception, then goes to prediction, and then goes to planning.
The one that by far is the most problematic is prediction. It’s not prediction of other automated cars, because if all cars were automated, this problem would be much more simple. How do you predict what a human being is going to do? That’s difficult for deep learning to learn right now.
Spectrum: Can you offset the weakness in prediction with stupendous perception?
Photo: Toyota Research Institute for Burgard
Wolfram Burgard
Wolfram Burgard: Yes, that is what car companies basically do. A camera provides semantics, lidar provides distance, radar provides velocities. But all this comes with problems, because sometimes you look at the world from different positions—that’s called parallax. Sometimes you don’t know which range estimate that pixel belongs to. That might make the decision complicated as to whether that is a person painted onto the side of a truck or whether this is an actual person.
With deep learning there is this promise that if you throw enough data at these networks, it’s going to work—finally. But it turns out that the amount of data that you need for self-driving cars is far larger than we expected.
Spectrum: When do deep learning’s limitations become apparent?
Pratt: The way to think about deep learning is that it’s really high-performance pattern matching. You have input and output as training pairs; you say this image should lead to that result; and you just do that again and again, for hundreds of thousands, millions of times.
Here’s the logical fallacy that I think most people have fallen prey to with deep learning. A lot of what we do with our brains can be thought of as pattern matching: “Oh, I see this stop sign, so I should stop.” But it doesn’t mean all of intelligence can be done through pattern matching.
“I asked myself, if all of those cars had automated drive, how good would they have to be to tolerate the number of crashes that would still occur?”
—Gill Pratt, Toyota Research Institute
For instance, when I’m driving and I see a mother holding the hand of a child on a corner and trying to cross the street, I am pretty sure she’s not going to cross at a red light and jaywalk. I know from my experience being a human being that mothers and children don’t act that way. On the other hand, say there are two teenagers—with blue hair, skateboards, and a disaffected look. Are they going to jaywalk? I look at that, you look at that, and instantly the probability in your mind that they’ll jaywalk is much higher than for the mother holding the hand of the child. It’s not that you’ve seen 100,000 cases of young kids—it’s that you understand what it is to be either a teenager or a mother holding a child’s hand.
You can try to fake that kind of intelligence. If you specifically train a neural network on data like that, you could pattern-match that. But you’d have to know to do it.
Spectrum: So you’re saying that when you substitute pattern recognition for reasoning, the marginal return on the investment falls off pretty fast?
Pratt: That’s absolutely right. Unfortunately, we don’t have the ability to make an AI that thinks yet, so we don’t know what to do. We keep trying to use the deep-learning hammer to hammer more nails—we say, well, let’s just pour more data in, and more data.
Spectrum: Couldn’t you train the deep-learning system to recognize teenagers and to assign the category a high propensity for jaywalking?
Burgard: People have been doing that. But it turns out that these heuristics you come up with are extremely hard to tweak. Also, sometimes the heuristics are contradictory, which makes it extremely hard to design these expert systems based on rules. This is where the strength of the deep-learning methods lies, because somehow they encode a way to see a pattern where, for example, here’s a feature and over there is another feature; it’s about the sheer number of parameters you have available.
Our separation of the components of a self-driving AI eases the development and even the learning of the AI systems. Some companies even think about using deep learning to do the job fully, from end to end, not having any structure at all—basically, directly mapping perceptions to actions.
Pratt: There are companies that have tried it; Nvidia certainly tried it. In general, it’s been found not to work very well. So people divide the problem into blocks, where we understand what each block does, and we try to make each block work well. Some of the blocks end up more like the expert system we talked about, where we actually code things, and other blocks end up more like machine learning.
Spectrum: So, what’s next—what new technique is in the offing?
Pratt: If I knew the answer, we’d do it. [Laughter]
Spectrum: You said that if all cars on the road were automated, the problem would be easy. Why not “geofence” the heck out of the self-driving problem, and have areas where only self-driving cars are allowed?
Pratt: That means putting in constraints on the operational design domain. This includes the geography—where the car should be automated; it includes the weather, it includes the level of traffic, it includes speed. If the car is going slow enough to avoid colliding without risking a rear-end collision, that makes the problem much easier. Street trolleys operate with traffic still in some parts of the world, and that seems to work out just fine. People learn that this vehicle may stop at unexpected times. My suspicion is, that is where we’ll see Level 4 autonomy in cities. It’s going to be in the lower speeds.
“We are now in the age of deep learning, and we don’t know what will come after.”
—Wolfram Burgard, Toyota Research Institute
That’s a sweet spot in the operational design domain, without a doubt. There’s another one at high speed on a highway, because access to highways is so limited. But unfortunately there is still the occasional debris that suddenly crosses the road, and the weather gets bad. The classic example is when somebody irresponsibly ties a mattress to the top of a car and it falls off; what are you going to do? And the answer is that terrible things happen—even for humans.
Spectrum: Learning by doing worked for the first cars, the first planes, the first steam boilers, and even the first nuclear reactors. We ran risks then; why not now?
Pratt: It has to do with the times. During the era where cars took off, all kinds of accidents happened, women died in childbirth, all sorts of diseases ran rampant; the expected characteristic of life was that bad things happened. Expectations have changed. Now the chance of dying in some freak accident is quite low because of all the learning that’s gone on, the OSHA [Occupational Safety and Health Administration] rules, UL code for electrical appliances, all the building standards, medicine.
Furthermore—and we think this is very important—we believe that empathy for a human being at the wheel is a significant factor in public acceptance when there is a crash. We don’t know this for sure—it’s a speculation on our part. I’ve driven, I’ve had close calls; that could have been me that made that mistake and had that wreck. I think people are more tolerant when somebody else makes mistakes, and there’s an awful crash. In the case of an automated car, we worry that that empathy won’t be there.
Photo: Toyota
Toyota is using this
Platform 4 automated driving test vehicle, based on the Lexus LS, to develop Level-4 self-driving capabilities for its “Chauffeur” project.
Spectrum: Toyota is building a system called Guardian to back up the driver, and a more futuristic system called Chauffeur, to replace the driver. How can Chauffeur ever succeed? It has to be better than a human plus Guardian!
Pratt: In the discussions we’ve had with others in this field, we’ve talked about that a lot. What is the standard? Is it a person in a basic car? Or is it a person with a car that has active safety systems in it? And what will people think is good enough?
These systems will never be perfect—there will always be some accidents, and no matter how hard we try there will still be occasions where there will be some fatalities. At what threshold are people willing to say that’s okay?
Spectrum: You were among the first top researchers to warn against hyping self-driving technology. What did you see that so many other players did not?
Pratt: First, in my own case, during my time at DARPA I worked on robotics, not cars. So I was somewhat of an outsider. I was looking at it from a fresh perspective, and that helps a lot.
Second, [when I joined Toyota in 2015] I was joining a company that is very careful—even though we have made some giant leaps—with the Prius hybrid drive system as an example. Even so, in general, the philosophy at Toyota is kaizen—making the cars incrementally better every single day. That care meant that I was tasked with thinking very deeply about this thing before making prognostications.
And the final part: It was a new job for me. The first night after I signed the contract I felt this incredible responsibility. I couldn’t sleep that whole night, so I started to multiply out the numbers, all using a factor of 10. How many cars do we have on the road? Cars on average last 10 years, though ours last 20, but let’s call it 10. They travel on an order of 10,000 miles per year. Multiply all that out and you get 10 to the 10th miles per year for our fleet on Planet Earth, a really big number. I asked myself, if all of those cars had automated drive, how good would they have to be to tolerate the number of crashes that would still occur? And the answer was so incredibly good that I knew it would take a long time. That was five years ago.
Burgard: We are now in the age of deep learning, and we don’t know what will come after. We are still making progress with existing techniques, and they look very promising. But the gradient is not as steep as it was a few years ago.
Pratt: There isn’t anything that’s telling us that it can’t be done; I should be very clear on that. Just because we don’t know how to do it doesn’t mean it can’t be done. Continue reading
Posted in Human Robots