Tag Archives: bad
#438798 This AI Thrashes the Hardest Atari Games ...
Learning from rewards seems like the simplest thing. I make coffee, I sip coffee, I’m happy. My brain registers “brewing coffee” as an action that leads to a reward.
That’s the guiding insight behind deep reinforcement learning, a family of algorithms that famously smashed most of Atari’s gaming catalog and triumphed over humans in strategy games like Go. Here, an AI “agent” explores the game, trying out different actions and registering ones that let it win.
Except it’s not that simple. “Brewing coffee” isn’t one action; it’s a series of actions spanning several minutes, where you’re only rewarded at the very end. By just tasting the final product, how do you learn to fine-tune grind coarseness, water to coffee ratio, brewing temperature, and a gazillion other factors that result in the reward—tasty, perk-me-up coffee?
That’s the problem with “sparse rewards,” which are ironically very abundant in our messy, complex world. We don’t immediately get feedback from our actions—no video-game-style dings or points for just grinding coffee beans—yet somehow we’re able to learn and perform an entire sequence of arm and hand movements while half-asleep.
This week, researchers from UberAI and OpenAI teamed up to bestow this talent on AI.
The trick is to encourage AI agents to “return” to a previous step, one that’s promising for a winning solution. The agent then keeps a record of that state, reloads it, and branches out again to intentionally explore other solutions that may have been left behind on the first go-around. Video gamers are likely familiar with this idea: live, die, reload a saved point, try something else, repeat for a perfect run-through.
The new family of algorithms, appropriately dubbed “Go-Explore,” smashed notoriously difficult Atari games like Montezuma’s Revenge that were previously unsolvable by its AI predecessors, while trouncing human performance along the way.
It’s not just games and digital fun. In a computer simulation of a robotic arm, the team found that installing Go-Explore as its “brain” allowed it to solve a challenging series of actions when given very sparse rewards. Because the overarching idea is so simple, the authors say, it can be adapted and expanded to other real-world problems, such as drug design or language learning.
Growing Pains
How do you reward an algorithm?
Rewards are very hard to craft, the authors say. Take the problem of asking a robot to go to a fridge. A sparse reward will only give the robot “happy points” if it reaches its destination, which is similar to asking a baby, with no concept of space and danger, to crawl through a potential minefield of toys and other obstacles towards a fridge.
“In practice, reinforcement learning works very well, if you have very rich feedback, if you can tell, ‘hey, this move is good, that move is bad, this move is good, that move is bad,’” said study author Joost Huinzinga. However, in situations that offer very little feedback, “rewards can intentionally lead to a dead end. Randomly exploring the space just doesn’t cut it.”
The other extreme is providing denser rewards. In the same robot-to-fridge example, you could frequently reward the bot as it goes along its journey, essentially helping “map out” the exact recipe to success. But that’s troubling as well. Over-holding an AI’s hand could result in an extremely rigid robot that ignores new additions to its path—a pet, for example—leading to dangerous situations. It’s a deceptive AI solution that seems effective in a simple environment, but crashes in the real world.
What we need are AI agents that can tackle both problems, the team said.
Intelligent Exploration
The key is to return to the past.
For AI, motivation usually comes from “exploring new or unusual situations,” said Huizinga. It’s efficient, but comes with significant downsides. For one, the AI agent could prematurely stop going back to promising areas because it thinks it had already found a good solution. For another, it could simply forget a previous decision point because of the mechanics of how it probes the next step in a problem.
For a complex task, the end result is an AI that randomly stumbles around towards a solution while ignoring potentially better ones.
“Detaching from a place that was previously visited after collecting a reward doesn’t work in difficult games, because you might leave out important clues,” Huinzinga explained.
Go-Explore solves these problems with a simple principle: first return, then explore. In essence, the algorithm saves different approaches it previously tried and loads promising save points—once more likely to lead to victory—to explore further.
Digging a bit deeper, the AI stores screen caps from a game. It then analyzes saved points and groups images that look alike as a potential promising “save point” to return to. Rinse and repeat. The AI tries to maximize its final score in the game, and updates its save points when it achieves a new record score. Because Atari doesn’t usually allow people to revisit any random point, the team used an emulator, which is a kind of software that mimics the Atari system but with custom abilities such as saving and reloading at any time.
The trick worked like magic. When pitted against 55 Atari games in the OpenAI gym, now commonly used to benchmark reinforcement learning algorithms, Go-Explore knocked out state-of-the-art AI competitors over 85 percent of the time.
It also crushed games previously unbeatable by AI. Montezuma’s Revenge, for example, requires you to move Pedro, the blocky protagonist, through a labyrinth of underground temples while evading obstacles such as traps and enemies and gathering jewels. One bad jump could derail the path to the next level. It’s a perfect example of sparse rewards: you need a series of good actions to get to the reward—advancing onward.
Go-Explore didn’t just beat all levels of the game, a first for AI. It also scored higher than any previous record for reinforcement learning algorithms at lower levels while toppling the human world record.
Outside a gaming environment, Go-Explore was also able to boost the performance of a simulated robot arm. While it’s easy for humans to follow high-level guidance like “put the cup on this shelf in a cupboard,” robots often need explicit training—from grasping the cup to recognizing a cupboard, moving towards it while avoiding obstacles, and learning motions to not smash the cup when putting it down.
Here, similar to the real world, the digital robot arm was only rewarded when it placed the cup onto the correct shelf, out of four possible shelves. When pitted against another algorithm, Go-Explore quickly figured out the movements needed to place the cup, while its competitor struggled with even reliably picking the cup up.
Combining Forces
By itself, the “first return, then explore” idea behind Go-Explore is already powerful. The team thinks it can do even better.
One idea is to change the mechanics of save points. Rather than reloading saved states through the emulator, it’s possible to train a neural network to do the same, without needing to relaunch a saved state. It’s a potential way to make the AI even smarter, the team said, because it can “learn” to overcome one obstacle once, instead of solving the same problem again and again. The downside? It’s much more computationally intensive.
Another idea is to combine Go-Explore with an alternative form of learning, called “imitation learning.” Here, an AI observes human behavior and mimics it through a series of actions. Combined with Go-Explore, said study author Adrien Ecoffet, this could make more robust robots capable of handling all the complexity and messiness in the real world.
To the team, the implications go far beyond Go-Explore. The concept of “first return, then explore” seems to be especially powerful, suggesting “it may be a fundamental feature of learning in general.” The team said, “Harnessing these insights…may be essential…to create generally intelligent agents.”
Image Credit: Adrien Ecoffet, Joost Huizinga, Joel Lehman, Kenneth O. Stanley, and Jeff Clune Continue reading
Posted in Human Robots
#438762 When Robots Enter the World, Who Is ...
Over the last half decade or so, the commercialization of autonomous robots that can operate outside of structured environments has dramatically increased. But this relatively new transition of robotic technologies from research projects to commercial products comes with its share of challenges, many of which relate to the rapidly increasing visibility that these robots have in society.
Whether it's because of their appearance of agency, or because of their history in popular culture, robots frequently inspire people’s imagination. Sometimes this is a good thing, like when it leads to innovative new use cases. And sometimes this is a bad thing, like when it leads to use cases that could be classified as irresponsible or unethical. Can the people selling robots do anything about the latter? And even if they can, should they?
Roboticists understand that robots, fundamentally, are tools. We build them, we program them, and even the autonomous ones are just following the instructions that we’ve coded into them. However, that same appearance of agency that makes robots so compelling means that it may not be clear to people without much experience with or exposure to real robots that a robot itself isn’t inherently good or bad—rather, as a tool, a robot is a reflection of its designers and users.
This can put robotics companies into a difficult position. When they sell a robot to someone, that person can, hypothetically, use the robot in any way they want. Of course, this is the case with every tool, but it’s the autonomous aspect that makes robots unique. I would argue that autonomy brings with it an implied association between a robot and its maker, or in this case, the company that develops and sells it. I’m not saying that this association is necessarily a reasonable one, but I think that it exists, even if that robot has been sold to someone else who has assumed full control over everything it does.
“All of our buyers, without exception, must agree that Spot will not be used to harm or intimidate people or animals, as a weapon or configured to hold a weapon”
—Robert Playter, Boston Dynamics
Robotics companies are certainly aware of this, because many of them are very careful about who they sell their robots to, and very explicit about what they want their robots to be doing. But once a robot is out in the wild, as it were, how far should that responsibility extend? And realistically, how far can it extend? Should robotics companies be held accountable for what their robots do in the world, or should we accept that once a robot is sold to someone else, responsibility is transferred as well? And what can be done if a robot is being used in an irresponsible or unethical way that could have a negative impact on the robotics community?
For perspective on this, we contacted folks from three different robotics companies, each of which has experience selling distinctive mobile robots to commercial end users. We asked them the same five questions about the responsibility that robotics companies have regarding the robots that they sell, and here’s what they had to say:
Do you have any restrictions on what people can do with your robots? If so, what are they, and if not, why not?
Péter Fankhauser, CEO, ANYbotics:
We closely work together with our customers to make sure that our solution provides the right approach for their problem. Thereby, the target use case is clear from the beginning and we do not work with customers interested in using our robot ANYmal outside the intended target applications. Specifically, we strictly exclude any military or weaponized uses and since the foundation of ANYbotics it is close to our heart to make human work easier, safer, and more enjoyable.
Robert Playter, CEO, Boston Dynamics:
Yes, we have restrictions on what people can do with our robots, which are outlined in our Terms and Conditions of Sale. All of our buyers, without exception, must agree that Spot will not be used to harm or intimidate people or animals, as a weapon or configured to hold a weapon. Spot, just like any product, must be used in compliance with the law.
Ryan Gariepy, CTO, Clearpath Robotics:
We do have strict restrictions and KYC processes which are based primarily on Canadian export control regulations. They depend on the type of equipment sold as well as where it is going. More generally, we also will not sell or support a robot if we know that it will create an uncontrolled safety hazard or if we have reason to believe that the buyer is unqualified to use the product. And, as always, we do not support using our products for the development of fully autonomous weapons systems.
More broadly, if you sell someone a robot, why should they be restricted in what they can do with it?
Péter Fankhauser, ANYbotics: We see the robot less as a simple object but more as an artificial workforce. This implies to us that the usage is closely coupled with the transfer of the robot and both the customer and the provider agree what the robot is expected to do. This approach is supported by what we hear from our customers with an increasing interest to pay for the robots as a service or per use.
Robert Playter, Boston Dynamics: We’re offering a product for sale. We’re going to do the best we can to stop bad actors from using our technology for harm, but we don’t have the control to regulate every use. That said, we believe that our business will be best served if our technology is used for peaceful purposes—to work alongside people as trusted assistants and remove them from harm’s way. We do not want to see our technology used to cause harm or promote violence. Our restrictions are similar to those of other manufacturers or technology companies that take steps to reduce or eliminate the violent or unlawful use of their products.
Ryan Gariepy, Clearpath Robotics: Assuming the organization doing the restricting is a private organization and the robot and its software is sold vs. leased or “managed,” there aren't strong legal reasons to restrict use. That being said, the manufacturer likewise has no obligation to continue supporting that specific robot or customer going forward. However, given that we are only at the very edge of how robots will reshape a great deal of society, it is in the best interest for the manufacturer and user to be honest with each other about their respective goals. Right now, you're not only investing in the initial purchase and relationship, you're investing in the promise of how you can help each other succeed in the future.
“If a robot is being used in a way that is irresponsible due to safety: intervene! If it’s unethical: speak up!”
—Péter Fankhauser, ANYbotics
What can you realistically do to make sure that people who buy your robots use them in the ways that you intend?
Péter Fankhauser, ANYbotics: We maintain a close collaboration with our customers to ensure their success with our solution. So for us, we have refrained from technical solutions to block unintended use.
Robert Playter, Boston Dynamics: We vet our customers to make sure that their desired applications are things that Spot can support, and are in alignment with our Terms and Conditions of Sale. We’ve turned away customers whose applications aren’t a good match with our technology. If customers misuse our technology, we’re clear in our Terms of Sale that their violations may void our warranty and prevent their robots from being updated, serviced, repaired, or replaced. We may also repossess robots that are not purchased, but leased. Finally, we will refuse future sales to customers that violate our Terms of Sale.
Ryan Gariepy, Clearpath Robotics: We typically work with our clients ahead of the purchase to make sure their expectations match reality, in particular on aspects like safety, supervisory requirements, and usability. It's far worse to sell a robot that'll sit on a shelf or worse, cause harm, then to not sell a robot at all, so we prefer to reduce the risk of this situation in advance of receiving an order or shipping a robot.
How do you evaluate the merit of edge cases, for example if someone wants to use your robot in research or art that may push the boundaries of what you personally think is responsible or ethical?
Péter Fankhauser, ANYbotics: It’s about the dialog, understanding, and figuring out alternatives that work for all involved parties and the earlier you can have this dialog the better.
Robert Playter, Boston Dynamics: There’s a clear line between exploring robots in research and art, and using the robot for violent or illegal purposes.
Ryan Gariepy, Clearpath Robotics: We have sold thousands of robots to hundreds of clients, and I do not recall the last situation that was not covered by a combination of export control and a general evaluation of the client's goals and expectations. I'm sure this will change as robots continue to drop in price and increase in flexibility and usability.
“You're not only investing in the initial purchase and relationship, you're investing in the promise of how you can help each other succeed in the future.”
—Ryan Gariepy, Clearpath Robotics
What should roboticists do if we see a robot being used in a way that we feel is unethical or irresponsible?
Péter Fankhauser, ANYbotics: If it’s irresponsible due to safety: intervene! If it’s unethical: speak up!
Robert Playter, Boston Dynamics: We want robots to be beneficial for humanity, which includes the notion of not causing harm. As an industry, we think robots will achieve long-term commercial viability only if people see robots as helpful, beneficial tools without worrying if they’re going to cause harm.
Ryan Gariepy, Clearpath Robotics: On a one off basis, they should speak to a combination of the user, the supplier or suppliers, the media, and, if safety is an immediate concern, regulatory or government agencies. If the situation in question risks becoming commonplace and is not being taken seriously, they should speak up more generally in appropriate forums—conferences, industry groups, standards bodies, and the like.
As more and more robots representing different capabilities become commercially available, these issues are likely to come up more frequently. The three companies we talked to certainly don’t represent every viewpoint, and we did reach out to other companies who declined to comment. But I would think (I would hope?) that everyone in the robotics community can agree that robots should be used in a way that makes people’s lives better. What “better” means in the context of art and research and even robots in the military may not always be easy to define, and inevitably there’ll be disagreement as to what is ethical and responsible, and what isn’t.
We’ll keep on talking about it, though, and do our best to help the robotics community to continue growing and evolving in a positive way. Let us know what you think in the comments. Continue reading
Posted in Human Robots
#438553 New Drone Software Handles Motor ...
Good as some drones are becoming at obstacle avoidance, accidents do still happen. And as far as robots go, drones are very much on the fragile side of things. Any sort of significant contact between a drone and almost anything else usually results in a catastrophic, out-of-control spin followed by a death plunge to the ground. Bad times. Bad, expensive times.
A few years ago, we saw some interesting research into software that can keep the most common drone form factor, the quadrotor, aloft and controllable even after the failure of one motor. The big caveat to that software was that it relied on GPS for state estimation, meaning that without a GPS signal, the drone is unable to get the information it needs to keep itself under control. In a paper recently accepted to RA-L, researchers at the University of Zurich report that they have developed a vision-based system that brings state estimation completely on-board. The upshot: potentially any drone with some software and a camera can keep itself safe even under the most challenging conditions.
A few years ago, we wrote about first author Sihao Sun’s work on high speed controlled flight of a quadrotor with a non-functional motor. But that innovation relied on an external motion capture system. Since then, Sun has moved from Tu Delft to Davide Scaramuzza’s lab at UZH, and it looks like he’s been able to combine his work on controlled spinning flight with the Robotics and Perception Group’s expertise in vision. Now, a downward-facing camera is all it takes for a spinning drone to remain stable and controllable:
Remember, this software isn’t just about guarding against motor failure. Drone motors themselves don’t just up and fail all that often, either with respect to their software or hardware. But they do represent the most likely point of failure for any drone, usually because when you run into something, what ultimately causes your drone to crash is damage to a motor or a propeller that causes loss of control.
The reason that earlier solutions relied on GPS was because the spinning drone needs a method of state estimation—that is, in order to be closed-loop controllable, the drone needs to have a reasonable understanding of what its position is and how that position is changing over time. GPS is an easy way to take care of this, but GPS is also an external system that doesn’t work everywhere. Having a state estimation system that’s completely internal to the drone itself is much more fail safe, and Sun got his onboard system to work through visual feature tracking with a downward-facing camera, even as the drone is spinning at over 20 rad/s.
While the system works well enough with a regular downward-facing camera—something that many consumer drones are equipped with for stabilization purposes—replacing it with an event camera (you remember event cameras, right?) makes the performance even better, especially in low light.
For more details on this, including what you’re supposed to do with a rapidly spinning partially disabled quadrotor (as well as what it’ll take to make this a standard feature on consumer hardware), we spoke with Sihao Sun via email.
IEEE Spectrum: what usually happens when a drone spinning this fast lands? Is there any way to do it safely?
Sihao Sun: Our experience shows that we can safely land the drone while it is spinning. When the range sensor measurements are lower than a threshold (around 10 cm, indicating that the drone is close to the ground), we switch off the rotors. During the landing procedure, despite the fast spinning motion, the thrust direction oscillates around the gravity vector, thus the drone touches the ground with its legs without damaging other components.
Can your system handle more than one motor failure?
Yes, the system can also handle the failure of two opposing rotors. However, if two adjacent rotors or more than two rotors fail, our method cannot save the quadrotor. Some research has shown that it is possible to control a quadrotor with only one remaining rotor. But the drone requires a very special inertial property, which is hard to satisfy in real applications.
How different is your system's performance from a similar system that relies on GPS, in a favorable environment?
In a favorable environment, our system outperforms those relying on GPS signals because it obtains better position estimates. Since a damaged quadrotor spins fast, the accelerometer readings are largely affected by centrifugal forces. When the GPS signal is lost or degraded, a drone relying on GPS needs to integrate these biased accelerometer measurements for position estimation, leading to large position estimation errors. Feeding these erroneous estimates to the flight controller can easily crash the drone.
When you say that your solution requires “only onboard sensors and computation,” are those requirements specialized, or would they be generally compatible with the current generation of recreational and commercial quadrotors?
We use an NVIDIA Jetson TX2 to run our solution, which includes two parts: the control algorithm and the vision-based state estimation algorithm. The control algorithm is lightweight; thus, we believe that it is compatible with the current generation of quadrotors. On the other hand, the vision-based state estimation requires relatively more computational resources, which may not be affordable for cheap recreational platforms. But this is not an issue for commercial quadrotors because many of them have more powerful processors than a TX2.
What else can event cameras be used for, in recreational or commercial applications?
Many drone applications can benefit from event cameras, especially those in high-speed or low-light conditions, such as autonomous drone racing, cave exploration, drone delivery during night time, etc. Event cameras also consume very little power, which is a significant advantage for energy-critical missions, such as planetary aerial vehicles for Mars explorations. Regarding space applications, we are currently collaborating with JPL to explore the use of event cameras to address the key limitations of standard cameras for the next Mars helicopter.
[ UZH RPG ] Continue reading
Posted in Human Robots
#437935 Start the New Year Right: By Watching ...
I don’t need to tell you that 2020 was a tough year. There was almost nothing good about it, and we saw it off with a “good riddance” and hopes for a better 2021. But robotics company Boston Dynamics took a different approach to closing out the year: when all else fails, why not dance?
The company released a video last week that I dare you to watch without laughing—or at the very least, cracking a pretty big smile. Because, well, dancing robots are funny. And it’s not just one dancing robot, it’s four of them: two humanoid Atlas bots, one four-legged Spot, and one Handle, a bot-on-wheels built for materials handling.
The robots’ killer moves look almost too smooth and coordinated to be real, leading many to speculate that the video was computer-generated. But if you can trust Elon Musk, there’s no CGI here.
This is not CGI https://t.co/VOivE97vPR
— Elon Musk (@elonmusk) December 29, 2020
Boston Dynamics went through a lot of changes in the last ten years; it was acquired by Google in 2013, then sold to Japanese conglomerate SoftBank in 2017 before being acquired again by Hyundai just a few weeks ago for $1.1 billion. But this isn’t the first time they teach a robot to dance and make a video for all the world to enjoy; Spot tore up the floor to “Uptown Funk” back in 2018.
Four-legged Spot went commercial in June, with a hefty price tag of $74,500, and was put to some innovative pandemic-related uses, including remotely measuring patients’ vital signs and reminding people to social distance.
Hyundai plans to implement its newly-acquired robotics prowess for everything from service and logistics robots to autonomous driving and smart factories.
They’ll have their work cut out for them. Besides being hilarious, kind of heartwarming, and kind of creepy all at once, the robots’ new routine is pretty impressive from an engineering standpoint. Compare it to a 2016 video of Atlas trying to pick up a box (I know it’s a machine with no feelings, but it’s hard not to feel a little bit bad for it, isn’t it?), and it’s clear Boston Dynamics’ technology has made huge strides. It wouldn’t be surprising if, in two years’ time, we see a video of a flash mob of robots whose routine includes partner dancing and back flips (which, admittedly, Atlas can already do).
In the meantime, though, this one is pretty entertaining—and not a bad note on which to start the new year.
Image Credit: Boston Dynamics Continue reading
Posted in Human Robots