Tag Archives: Machine Learning
#436021 AI Faces Speed Bumps and Potholes on Its ...
Implementing machine learning in the real world isn’t easy. The tools are available and the road is well-marked—but the speed bumps are many.
That was the conclusion of panelists wrapping up a day of discussions at the IEEE AI Symposium 2019, held at Cisco’s San Jose, Calif., campus last week.
The toughest problem, says Ben Irving, senior manager of Cisco’s strategy innovations group, is people.
It’s tough to find data scientist expertise, he indicated, so companies are looking into non-traditional sources of personnel, like political science. “There are some untapped areas with a lot of untapped data science expertise,” Irving says.
Lazard’s artificial intelligence manager Trevor Mottl agreed that would-be data scientists don’t need formal training or experience to break into the field. “This field is changing really rapidly,” he says. “There are new language models coming out every month, and new tools, so [anyone should] expect to not know everything. Experiment, try out new tools and techniques, read, study, spend time; there aren’t any true experts at this point because the foundational elements are shifting so rapidly.”
“It is a wonderful time to get into a field,” he reasons, noting that it doesn’t take long to catch up because there aren’t 20 years of history.”
Confusion about what different kinds of machine learning specialists do doesn’t help the personnel situation. An audience member asked panelists to explain the difference between data scientist, data analyst, and data engineer. Darrin Johnson, Nvidia global director of technical marketing for enterprise, admitted it’s hard to sort out, and any two companies could define the positions differently. “Sometimes,” he says, particularly at smaller companies, “a data scientist plays all three roles. But as companies grow, there are different groups that ingest data, clean data, and use data. At some companies, training and inference are separate. It really depends, which is a challenge when you are trying to hire someone.”
Mitigating the risks of a hot job market
The competition to hire data scientists, analysts, engineers, or whatever companies call them requires that managers make sure any work being done is structured and comprehensible at all times, the panelists cautioned.
“We need to remember that our data scientists go home every day and sometimes they don’t come back because they go home and then go to a different company,” says Lazard’s Mottl. “That’s a fact of life. If you give people choice on [how they do development], and have a successful person who gets poached by competitor, you have to either hire a team to unwrap what that person built or jettison their work and rebuild it.”
By contrast, he says, “places that have structured coding and structured commits and organized constructions of software have done very well.”
But keeping all of a company’s engineers working with the same languages and on the same development paths is not easy to do in a field that moves as fast as machine learning. Zongjie Diao, Cisco director of product management for machine learning, quipped: “I have a data scientist friend who says the speed at which he changes girlfriends is less than speed at which he changes languages.”
The data scientist/IT manager clash
Once a company finds the data engineers and scientists they need and get them started on the task of applying machine learning to that company’s operations, one of the first obstacles they face just might be the company’s IT department, the panelists suggested.
“IT is process oriented,” Mottl says. The IT team “knows how to keep data secure, to set up servers. But when you bring in a data science team, they want sandboxes, they want freedom, they want to explore and play.”
Also, Nvidia’s Johnson pointed out, “There is a language barrier.” The AI world, he says, is very different from networking or storage, and data scientists find it hard to articulate their requirements to IT.
On the ground or in the cloud?
And then there is the decision of where exactly machine learning should happen—on site, or in the cloud? At Lazard, Mottl says, the deep learning engineers do their experimentation on premises; that’s their sandbox. “But when we deploy, we deploy in the cloud,” he says.
Nvidia, Johnson says, thinks the opposite approach is better. We see the cloud as “the sandbox,” he says. “So you can run as many experiments as possible, fail fast, and learn faster.”
For Cisco’s Irving, the “where” of machine learning depends on the confidentiality of the data.
Mottl, who says rolling machine learning technology into operation can hit resistance from all across the company, had one last word of caution for those aiming to implement AI:
Data scientists are building things that might change the ways other people in the organization work, like sales and even knowledge workers. [You need to] think about the internal stakeholders and prepare them, because the last thing you want to do is to create a valuable new thing that nobody likes and people take potshots against.
The AI Symposium was organized by the Silicon Valley chapters of the IEEE Young Professionals, the IEEE Consultants’ Network, and IEEE Women in Engineering and supported by Cisco. Continue reading
Posted in Human Robots
#435824 A Q&A with Cruise’s head of AI, ...
In 2016, Cruise, an autonomous vehicle startup acquired by General Motors, had about 50 employees. At the beginning of 2019, the headcount at its San Francisco headquarters—mostly software engineers, mostly working on projects connected to machine learning and artificial intelligence—hit around 1000. Now that number is up to 1500, and by the end of this year it’s expected to reach about 2000, sprawling into a recently purchased building that had housed Dropbox. And that’s not counting the 200 or so tech workers that Cruise is aiming to install in a Seattle, Wash., satellite development center and a handful of others in Phoenix, Ariz., and Pasadena, Calif.
Cruise’s recent hires aren’t all engineers—it takes more than engineering talent to manage operations. And there are hundreds of so-called safety drivers that are required to sit in the 180 or so autonomous test vehicles whenever they roam the San Francisco streets. But that’s still a lot of AI experts to be hiring in a time of AI engineer shortages.
Hussein Mehanna, head of AI/ML at Cruise, says the company’s hiring efforts are on track, due to the appeal of the challenge of autonomous vehicles in drawing in AI experts from other fields. Mehanna himself joined Cruise in May from Google, where he was director of engineering at Google Cloud AI. Mehanna had been there about a year and a half, a relatively quick career stop after a short stint at Snap following four years working in machine learning at Facebook.
Mehanna has been immersed in AI and machine learning research since his graduate studies in speech recognition and natural language processing at the University of Cambridge. I sat down with Mehanna to talk about his career, the challenges of recruiting AI experts and autonomous vehicle development in general—and some of the challenges specific to San Francisco. We were joined by Michael Thomas, Cruise’s manager of AI/ML recruiting, who had also spent time recruiting AI engineers at Google and then Facebook.
IEEE Spectrum: When you were at Cambridge, did you think AI was going to take off like a rocket?
Mehanna: Did I imagine that AI was going to be as dominant and prevailing and sometimes hyped as it is now? No. I do recall in 2003 that my supervisor and I were wondering if neural networks could help at all in speech recognition. I remember my supervisor saying if anyone could figure out how use a neural net for speech he would give them a grant immediately. So he was on the right path. Now neural networks have dominated vision, speech, and language [processing]. But that boom started in 2012.
“In the early days, Facebook wasn’t that open to PhDs, it actually had a negative sentiment about researchers, and then Facebook shifted”
I didn’t [expect it], but I certainly aimed for it when [I was at] Microsoft, where I deliberately pushed my career towards machine learning instead of big data, which was more popular at the time. And [I aimed for it] when I joined Facebook.
In the early days, Facebook wasn’t that open to PhDs, or researchers. It actually had a negative sentiment about researchers. And then Facebook shifted to becoming one of the key places where PhD students wanted to do internships or join after they graduated. It was a mindset shift, they were [once] at a point in time where they thought what was needed for success wasn’t research, but now it’s different.
There was definitely an element of risk [in taking a machine learning career path], but I was very lucky, things developed very fast.
IEEE Spectrum: Is it getting harder or easier to find AI engineers to hire, given the reported shortages?
Mehanna: There is a mismatch [between job openings and qualified engineers], though it is hard to quantify it with numbers. There is good news as well: I see a lot more students diving deep into machine learning and data in their [undergraduate] computer science studies, so it’s not as bleak as it seems. But there is massive demand in the market.
Here at Cruise, demand for AI talent is just growing and growing. It might be is saturating or slowing down at other kinds of companies, though, [which] are leveraging more traditional applications—ad prediction, recommendations—that have been out there in the market for a while. These are more mature, better understood problems.
I believe autonomous vehicle technologies is the most difficult AI problem out there. The magnitude of the challenge of these problems is 1000 times more than other problems. They aren’t as well understood yet, and they require far deeper technology. And also the quality at which they are expected to operate is off the roof.
The autonomous vehicle problem is the engineering challenge of our generation. There’s a lot of code to write, and if we think we are going to hire armies of people to write it line by line, it’s not going to work. Machine learning can accelerate the process of generating the code, but that doesn’t mean we aren’t going to have engineers; we actually need a lot more engineers.
Sometimes people worry that AI is taking jobs. It is taking some developer jobs, but it is actually generating other developer jobs as well, protecting developers from the mundane and helping them build software faster and faster.
IEEE Spectrum: Are you concerned that the demand for AI in industry is drawing out the people in academia who are needed to educate future engineers, that is, the “eating the seed corn” problem?
Mehanna: There are some negative examples in the industry, but that’s not our style. We are looking for collaborations with professors, we want to cultivate a very deep and respectful relationship with universities.
And there’s another angle to this: Universities require a thriving industry for them to thrive. It is going to be extremely beneficial for academia to have this flourishing industry in AI, because it attracts more students to academia. I think we are doing them a fantastic favor by building these career opportunities. This is not the same as in my early days, [when] people told me “don’t go to AI; go to networking, work in the mobile industry; mobile is flourishing.”
IEEE Spectrum: Where are you looking as you try to find a thousand or so engineers to hire this year?
Thomas: We look for people who want to use machine learning to solve problems. They can be in many different industries—in the financial markets, in social media, in advertising. The autonomous vehicle industry is in its infancy. You can compare it to mobile in the early days: When the iPhone first came out, everyone was looking for developers with mobile experience, but you weren’t going to find them unless you went to straight to Apple, [so you had to hire other kinds of engineers]. This is the same type of thing: it is so new that you aren’t going to find experts in this area, because we are all still learning.
“You don’t have to be an autonomous vehicle expert to flourish in this world. It’s not too late to move…now would be a great time for AI experts working on other problems to shift their attention to autonomous vehicles.”
Mehanna: Because autonomous vehicle technology is the new frontier for AI experts, [the number of] people with both AI and autonomous vehicle experience is quite limited. So we are acquiring AI experts wherever they are, and helping them grow into the autonomous vehicle area. You don’t have to be an autonomous vehicle expert to flourish in this world. It’s not too late to move; even though there is a lot of great tech developed, there’s even more innovation ahead, so now would be a great time for AI experts working on other problems or applications to shift their attention to autonomous vehicles.
It feels like the Internet in 1980. It’s about to happen, but there are endless applications [to be developed over] the next few decades. Even if we can get a car to drive safely, there is the question of how can we tune the ride comfort, and then applying it all to different cities, different vehicles, different driving situations, and who knows to what other applications.
I can see how I can spend a lifetime career trying to solve this problem.
IEEE Spectrum: Why are you doing most of your development in San Francisco?
Mehanna: I think the best talent of the world is in Silicon Valley, and solving the autonomous vehicle problem is going to require the best of the best. It’s not just the engineering talent that is here, but [also] the entrepreneurial spirit. Solving the problem just as a technology is not going to be successful, you need to solve the product and the technology together. And the entrepreneurial spirit is one of the key reasons Cruise secured 7.5 billion in funding [besides GM, the company has a number of outside investors, including Honda, Softbank, and T. Rowe Price]. That [funding] is another reason Cruise is ahead of many others, because this problem requires deep resources.
“If you can do an autonomous vehicle in San Francisco you can do it almost anywhere.”
[And then there is the driving environment.] When I speak to my peers in the industry, they have a lot of respect for us, because the problems to solve in San Francisco technically are an order of magnitude harder. It is a tight environment, with a lot of pedestrians, and driving patterns that, let’s put it this way, are not necessarily the best in the nation. Which means we are seeing more problems ahead of our competitors, which gets us to better [software]. I think if you can do an autonomous vehicle in San Francisco you can do it almost anywhere.
A version of this post appears in the September 2019 print magazine as “AI Engineers: The Autonomous-Vehicle Industry Wants You.” Continue reading
Posted in Human Robots
#435775 Jaco Is a Low-Power Robot Arm That Hooks ...
We usually think of robots as taking the place of humans in various tasks, but robots of all kinds can also enhance human capabilities. This may be especially true for people with disabilities. And while the Cybathlon competition showed what's possible when cutting-edge research robotics is paired with expert humans, that competition isn't necessarily reflective of the kind of robotics available to most people today.
Kinova Robotics's Jaco arm is an assistive robotic arm designed to be mounted on an electric wheelchair. With six degrees of freedom plus a three-fingered gripper, the lightweight carbon fiber arm is frequently used in research because it's rugged and versatile. But from the start, Kinova created it to add autonomy to the lives of people with mobility constraints.
Earlier this year, Kinova shared the story of Mary Nelson, an 11-year-old girl with spinal muscular atrophy, who uses her Jaco arm to show her horse in competition. Spinal muscular atrophy is a neuromuscular disorder that impairs voluntary muscle movement, including muscles that help with respiration, and Mary depends on a power chair for mobility.
We wanted to learn more about how Kinova designs its Jaco arm, and what that means for folks like Mary, so we spoke with both Kinova and Mary's parents to find out how much of a difference a robot arm can make.
IEEE Spectrum: How did Mary interact with the world before having her arm, and what was involved in the decision to try a robot arm in general? And why then Kinova's arm specifically?
Ryan Nelson: Mary interacts with the world much like you and I do, she just uses different tools to do so. For example, she is 100 percent independent using her computer, iPad, and phone, and she prefers to use a mouse. However, she cannot move a standard mouse, so she connects her wheelchair to each device with Bluetooth to move the mouse pointer/cursor using her wheelchair joystick.
For years, we had a Manfrotto magic arm and super clamp attached to her wheelchair and she used that much like the robotic arm. We could put a baseball bat, paint brush, toys, etc. in the super clamp so that Mary could hold the object and interact as physically able children do. Mary has always wanted to be more independent, so we knew the robotic arm was something she must try. We had seen videos of the Kinova arm on YouTube and on their website, so we reached out to them to get a trial.
Can you tell us about the Jaco arm, and how the process of designing an assistive robot arm is different from the process of designing a conventional robot arm?
Nathaniel Swenson, Director of U.S. Operations — Assistive Technologies at Kinova: Jaco is our flagship robotic arm. Inspired by our CEO's uncle and its namesake, Jacques “Jaco” Forest, it was designed as assistive technology with power wheelchair users in mind.
The primary differences between Jaco and our other robots, such as the new Gen3, which was designed to meet the needs of academic and industry research teams, are speed and power consumption. Other robots such as the Gen3 can move faster and draw slightly more power because they aren't limited by the battery size of power wheelchairs. Depending on the use case, they might not interact directly with a human being in the research setting and can safely move more quickly. Jaco is designed to move at safe speeds and make direct contact with the end user and draw very little power directly from their wheelchair.
The most important consideration in the design process of an assistive robot is the safety of the end user. Jaco users operate their robots through their existing drive controls to assist them in daily activities such as eating, drinking, and opening doors and they don't have to worry about the robot draining their chair's batteries throughout the day. The elegant design that results from meeting the needs of our power chair users has benefited subsequent iterations, [of products] such as the Gen3, as well: Kinova's robots are lightweight, extremely efficient in their power consumption, and safe for direct human-robot interaction. This is not true of conventional industrial robots.
What was the learning process like for Mary? Does she feel like she's mastered the arm, or is it a continuous learning process?
Ryan Nelson: The learning process was super quick for Mary. However, she amazes us every day with the new things that she can do with the arm. Literally within minutes of installing the arm on her chair, Mary had it figured out and was shaking hands with the Kinova rep. The control of the arm is super intuitive and the Kinova reps say that SMA (Spinal Muscular Atrophy) children are perfect users because they are so smart—they pick it up right away. Mary has learned to do many fine motor tasks with the arm, from picking up small objects like a pencil or a ruler, to adjusting her glasses on her face, to doing science experiments.
Photo: The Nelson Family
Mary uses a headset microphone to amplify her voice, and she will use the arm and finger to adjust the microphone in front of her mouth after she is done eating (also a task she mastered quickly with the arm). Additionally, Mary will use the arms to reach down and adjust her feet or leg by grabbing them with the arm and moving them to a more comfortable position. All of these examples are things she never really asked us to do, but something she needed and just did on her own, with the help of the arm.
What is the most common feedback that you get from new users of the arm? How about from experienced users who have been using the arm for a while?
Nathaniel Swenson: New users always tell us how excited they are to see what they can accomplish with their new Jaco. From day one, they are able to do things that they have longed to do without assistance from a caregiver: take a drink of water or coffee, scratch an itch, push the button to open an “accessible” door or elevator, or even feed their baby with a bottle.
The most common feedback I hear from experienced users is that Jaco has changed their life. Our experienced users like Mary are rock stars: everywhere they go, people get excited to see what they'll do next. The difference between a new user and an experienced user could be as little as two weeks. People who operate power wheelchairs every day are already expert drivers and we just add a new “gear” to their chair: robot mode. It's fun to see how quickly new users master the intuitive Jaco control modes.
What changes would you like to see in the next generation of Jaco arm?
Ryan Nelson: Titanium fingers! Make it lift heavier objects, hold heavier items like a baseball bat, machine gun, flame thrower, etc., and Mary literally said this last night: “I wish the arm moved fast enough to play the piano.”
Nathaniel Swenson: I love the idea of titanium fingers! Jaco's fingers are made from a flexible polymer and designed to avoid harm. This allows the fingers to bend or dislocate, rather than break, but it also means they are not as durable as a material like titanium. Increased payload, the ability to manipulate heavier objects, requires increased power consumption. We've struck a careful balance between providing enough strength to accomplish most medically necessary Activities of Daily Living and efficient use of the power chair's batteries.
We take Isaac Asimov's Laws of Robotics pretty seriously. When we start to combine machine guns, flame throwers, and artificial intelligence with robots, I get very nervous!
I wish the arm moved fast enough to play the piano, too! I am also a musician and I share Mary's dream of an assistive robot that would enable her to make music. In the meantime, while we work on that, please enjoy this beautiful violin piece by Manami Ito and her one-of-a-kind violin prosthesis:
To what extent could more autonomy for the arm be helpful for users? What would be involved in implementing that?
Nathaniel Swenson: Artificial intelligence, machine learning, and deep learning will introduce greater autonomy in future iterations of assistive robots. This will enable them to perform more complex tasks that aren't currently possible, and enable them to accomplish routine tasks more quickly and with less input than the current manual control requires.
For assistive robots, implementation of greater autonomy involves a focus on end-user safety and improvements in the robot's awareness of its environment. Autonomous robots that work in close proximity with humans need vision. They must be able to see to avoid collisions and they use haptic feedback to tell the robot how much force is being exerted on objects. All of these technologies exist, but the largest obstacle to bringing them to the assistive technology market is to prove to the health insurance companies who will fund them that they are both safe and medically necessary. Continue reading
Posted in Human Robots
#435748 Video Friday: This Robot Is Like a ...
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!):
RSS 2019 – June 22-26, 2019 – Freiburg, Germany
Hamlyn Symposium on Medical Robotics – June 23-26, 2019 – London, U.K.
ETH Robotics Summer School – June 27-1, 2019 – Zurich, Switzerland
MARSS 2019 – July 1-5, 2019 – Helsinki, Finland
ICRES 2019 – July 29-30, 2019 – London, U.K.
DARPA SubT Tunnel Circuit – August 15-22, 2019 – Pittsburgh, Pa., USA
Let us know if you have suggestions for next week, and enjoy today’s videos.
It’s been a while since we last spoke to Joe Jones, the inventor of Roomba, about his solar-powered, weed-killing robot, called Tertill, which he was launching as a Kickstarter project. Tertill is now available for purchase (US $300) and is shipping right now.
[ Tertill ]
Usually, we don’t post videos that involve drone use that looks to be either illegal or unsafe. These flights over the protests in Hong Kong are almost certainly both. However, it’s also a unique perspective on the scale of these protests.
[ Team BlackSheep ]
ICYMI: iRobot announced this week that it has acquired Root Robotics.
[ iRobot ]
This Boston Dynamics parody video went viral this week.
The CGI is good but the gratuitous violence—even if it’s against a fake robot—is a bit too much?
This is still our favorite Boston Dynamics parody video:
[ Corridor ]
Biomedical Engineering Department Head Bin He and his team have developed the first-ever successful non-invasive mind-controlled robotic arm to continuously track a computer cursor.
[ CMU ]
Organic chemists, prepare to meet your replacement:
Automated chemical synthesis carries great promises of safety, efficiency and reproducibility for both research and industry laboratories. Current approaches are based on specifically-designed automation systems, which present two major drawbacks: (i) existing apparatus must be modified to be integrated into the automation systems; (ii) such systems are not flexible and would require substantial re-design to handle new reactions or procedures. In this paper, we propose a system based on a robot arm which, by mimicking the motions of human chemists, is able to perform complex chemical reactions without any modifications to the existing setup used by humans. The system is capable of precise liquid handling, mixing, filtering, and is flexible: new skills and procedures could be added with minimum effort. We show that the robot is able to perform a Michael reaction, reaching a yield of 34%, which is comparable to that obtained by a junior chemist (undergraduate student in Chemistry).
[ arXiv ] via [ NTU ]
So yeah, ICRA 2019 was huge and awesome. Here are some brief highlights.
[ Montreal Gazette ]
For about US $5, this drone will deliver raw meat and beer to you if you live on an uninhabited island in Tokyo Bay.
[ Nikkei ]
The Smart Microsystems Lab at Michigan State University has a new version of their Autonomous Surface Craft. It’s autonomous, open source, and awfully hard to sink.
[ SML ]
As drone shows go, this one is pretty good.
[ CCTV ]
Here’s a remote controlled robot shooting stuff with a very large gun.
[ HDT ]
Over a period of three quarters (September 2018 thru May 2019), we’ve had the opportunity to work with five graduating University of Denver students as they brought their idea for a Misty II arm extension to life.
[ Misty Robotics ]
If you wonder how it looks to inspect burners and superheaters of a boiler with an Elios 2, here you are! This inspection was performed by Svenska Elektrod in a peat-fired boiler for Vattenfall in Sweden. Enjoy!
[ Flyability ]
The newest Soft Robotics technology, mGrip mini fingers, made for tight spaces, small packaging, and delicate items, giving limitless opportunities for your applications.
[ Soft Robotics ]
What if legged robots were able to generate dynamic motions in real-time while interacting with a complex environment? Such technology would represent a significant step forward the deployment of legged systems in real world scenarios. This means being able to replace humans in the execution of dangerous tasks and to collaborate with them in industrial applications.
This workshop aims to bring together researchers from all the relevant communities in legged locomotion such as: numerical optimization, machine learning (ML), model predictive control (MPC) and computational geometry in order to chart the most promising methods to address the above-mentioned scientific challenges.
[ Num Opt Wkshp ]
Army researchers teamed with the U.S. Marine Corps to fly and test 3-D printed quadcopter prototypes a the Marine Corps Air Ground Combat Center in 29 Palms, California recently.
[ CCDC ARL ]
Lex Fridman’s Artificial Intelligence podcast featuring Rosalind Picard.
[ AI Podcast ]
In this week’s episode of Robots in Depth, per speaks with Christian Guttmann, executive director of the Nordic AI Artificial Intelligence Institute.
Christian Guttmann talks about AI and wanting to understand intelligence enough to recreate it. Christian has be focusing on AI in healthcare and has recently started to communicate the opportunities and challenges in artificial intelligence to the general public. This is something that the host Per Sjöborg is also very passionate about. We also get to hear about the Nordic AI institute and the work it does to inform all parts of society about AI.
[ Robots in Depth ] Continue reading
Posted in Human Robots
#435707 AI Agents Startle Researchers With ...
After 25 million games, the AI agents playing hide-and-seek with each other had mastered four basic game strategies. The researchers expected that part.
After a total of 380 million games, the AI players developed strategies that the researchers didn’t know were possible in the game environment—which the researchers had themselves created. That was the part that surprised the team at OpenAI, a research company based in San Francisco.
The AI players learned everything via a machine learning technique known as reinforcement learning. In this learning method, AI agents start out by taking random actions. Sometimes those random actions produce desired results, which earn them rewards. Via trial-and-error on a massive scale, they can learn sophisticated strategies.
In the context of games, this process can be abetted by having the AI play against another version of itself, ensuring that the opponents will be evenly matched. It also locks the AI into a process of one-upmanship, where any new strategy that emerges forces the opponent to search for a countermeasure. Over time, this “self-play” amounted to what the researchers call an “auto-curriculum.”
According to OpenAI researcher Igor Mordatch, this experiment shows that self-play “is enough for the agents to learn surprising behaviors on their own—it’s like children playing with each other.”
Reinforcement is a hot field of AI research right now. OpenAI’s researchers used the technique when they trained a team of bots to play the video game Dota 2, which squashed a world-champion human team last April. The Alphabet subsidiary DeepMind has used it to triumph in the ancient board game Go and the video game StarCraft.
Aniruddha Kembhavi, a researcher at the Allen Institute for Artificial Intelligence (AI2) in Seattle, says games such as hide-and-seek offer a good way for AI agents to learn “foundational skills.” He worked on a team that taught their AllenAI to play Pictionary with humans, viewing the gameplay as a way for the AI to work on common sense reasoning and communication. “We are, however, quite far away from being able to translate these preliminary findings in highly simplified environments into the real world,” says Kembhavi.
Illustration: OpenAI
AI agents construct a fort during a hide-and-seek game developed by OpenAI.
In OpenAI’s game of hide-and-seek, both the hiders and the seekers received a reward only if they won the game, leaving the AI players to develop their own strategies. Within a simple 3D environment containing walls, blocks, and ramps, the players first learned to run around and chase each other (strategy 1). The hiders next learned to move the blocks around to build forts (2), and then the seekers learned to move the ramps (3), enabling them to jump inside the forts. Then the hiders learned to move all the ramps into their forts before the seekers could use them (4).
The two strategies that surprised the researchers came next. First the seekers learned that they could jump onto a box and “surf” it over to a fort (5), allowing them to jump in—a maneuver that the researchers hadn’t realized was physically possible in the game environment. So as a final countermeasure, the hiders learned to lock all the boxes into place (6) so they weren’t available for use as surfboards.
Illustration: OpenAI
An AI agent uses a nearby box to surf its way into a competitor’s fort.
In this circumstance, having AI agents behave in an unexpected way wasn’t a problem: They found different paths to their rewards, but didn’t cause any trouble. However, you can imagine situations in which the outcome would be rather serious. Robots acting in the real world could do real damage. And then there’s Nick Bostrom’s famous example of a paper clip factory run by an AI, whose goal is to make as many paper clips as possible. As Bostrom told IEEE Spectrum back in 2014, the AI might realize that “human bodies consist of atoms, and those atoms could be used to make some very nice paper clips.”
Bowen Baker, another member of the OpenAI research team, notes that it’s hard to predict all the ways an AI agent will act inside an environment—even a simple one. “Building these environments is hard,” he says. “The agents will come up with these unexpected behaviors, which will be a safety problem down the road when you put them in more complex environments.”
AI researcher Katja Hofmann at Microsoft Research Cambridge, in England, has seen a lot of gameplay by AI agents: She started a competition that uses Minecraft as the playing field. She says the emergent behavior seen in this game, and in prior experiments by other researchers, shows that games can be a useful for studies of safe and responsible AI.
“I find demonstrations like this, in games and game-like settings, a great way to explore the capabilities and limitations of existing approaches in a safe environment,” says Hofmann. “Results like these will help us develop a better understanding on how to validate and debug reinforcement learning systems–a crucial step on the path towards real-world applications.”
Baker says there’s also a hopeful takeaway from the surprises in the hide-and-seek experiment. “If you put these agents into a rich enough environment they will find strategies that we never knew were possible,” he says. “Maybe they can solve problems that we can’t imagine solutions to.” Continue reading
Posted in Human Robots