Tag Archives: taught
#436426 Video Friday: This Robot Refuses to Fall ...
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!):
Robotic Arena – January 25, 2020 – Wrocław, Poland
DARPA SubT Urban Circuit – February 18-27, 2020 – Olympia, Wash., USA
Let us know if you have suggestions for next week, and enjoy today’s videos.
In case you somehow missed the massive Skydio 2 review we posted earlier this week, the first batches of the drone are now shipping. Each drone gets a lot of attention before it goes out the door, and here’s a behind-the-scenes clip of the process.
[ Skydio ]
Sphero RVR is one of the 15 robots on our robot gift guide this year. Here’s a new video Sphero just released showing some of the things you can do with the robot.
[ RVR ]
NimbRo-OP2 has some impressive recovery skills from the obligatory research-motivated robot abuse.
[ NimbRo ]
Teams seeking to qualify for the Virtual Urban Circuit of the Subterranean Challenge can access practice worlds to test their approaches prior to submitting solutions for the competition. This video previews three of the practice environments.
[ DARPA SubT ]
Stretchable skin-like robots that can be rolled up and put in your pocket have been developed by a University of Bristol team using a new way of embedding artificial muscles and electrical adhesion into soft materials.
[ Bristol ]
Happy Holidays from ABB!
Helping New York celebrate the festive season, twelve ABB robots are interacting with visitors to Bloomingdale’s iconic holiday celebration at their 59th Street flagship store. ABB’s robots are the main attraction in three of Bloomingdale’s twelve-holiday window displays at Lexington and Third Avenue, as ABB demonstrates the potential for its robotics and automation technology to revolutionize visual merchandising and make the retail experience more dynamic and whimsical.
[ ABB ]
We introduce pelican eel–inspired dual-morphing architectures that embody quasi-sequential behaviors of origami unfolding and skin stretching in response to fluid pressure. In the proposed system, fluid paths were enclosed and guided by a set of entirely stretchable origami units that imitate the morphing principle of the pelican eel’s stretchable and foldable frames. This geometric and elastomeric design of fluid networks, in which fluid pressure acts in the direction that the whole body deploys first, resulted in a quasi-sequential dual-morphing response. To verify the effectiveness of our design rule, we built an artificial creature mimicking a pelican eel and reproduced biomimetic dual-morphing behavior.
And here’s a real pelican eel:
[ Science Robotics ]
Delft Dynamics’ updated anti-drone system involves a tether, mid-air net gun, and even a parachute.
[ Delft Dynamics ]
Teleoperation is a great way of helping robots with complex tasks, especially if you can do it through motion capture. But what if you’re teleoperating a non-anthropomorphic robot? Columbia’s ROAM Lab is working on it.
[ Paper ] via [ ROAM Lab ]
I don’t know how I missed this video last year because it’s got a steely robot hand squeezing a cute lil’ chick.
[ MotionLib ] via [ RobotStart ]
In this video we present results of a trajectory generation method for autonomous overtaking of unexpected obstacles in a dynamic urban environment. In these settings, blind spots can arise from perception limitations. For example when overtaking unexpected objects on the vehicle’s ego lane on a two-way street. In this case, a human driver would first make sure that the opposite lane is free and that there is enough room to successfully execute the maneuver, and then it would cut into the opposite lane in order to execute the maneuver successfully. We consider the practical problem of autonomous overtaking when the coverage of the perception system is impaired due to occlusion.
[ Paper ]
New weirdness from Toio!
[ Toio ]
Palo Alto City Library won a technology innovation award! Watch to see how Senior Librarian Dan Lou is using Misty to enhance their technology programs to inspire and educate customers.
[ Misty Robotics ]
We consider the problem of reorienting a rigid object with arbitrary known shape on a table using a two-finger pinch gripper. Reorienting problem is challenging because of its non-smoothness and high dimensionality. In this work, we focus on solving reorienting using pivoting, in which we allow the grasped object to rotate between fingers. Pivoting decouples the gripper rotation from the object motion, making it possible to reorient an object under strict robot workspace constraints.
[ CMU ]
How can a mobile robot be a good pedestrian without bumping into you on the sidewalk? It must be hard for a robot to navigate in crowded environments since the flow of traffic follows implied social rules. But researchers from MIT developed an algorithm that teaches mobile robots to maneuver in crowds of people, respecting their natural behaviour.
[ Roboy Research Reviews ]
What happens when humans and robots make art together? In this awe-inspiring talk, artist Sougwen Chung shows how she “taught” her artistic style to a machine — and shares the results of their collaboration after making an unexpected discovery: robots make mistakes, too. “Part of the beauty of human and machine systems is their inherent, shared fallibility,” she says.
[ TED ]
Last month at the Cooper Union in New York City, IEEE TechEthics hosted a public panel session on the facts and misperceptions of autonomous vehicles, part of the IEEE TechEthics Conversations Series. The speakers were: Jason Borenstein from Georgia Tech; Missy Cummings from Duke University; Jack Pokrzywa from SAE; and Heather M. Roff from Johns Hopkins Applied Physics Laboratory. The panel was moderated by Mark A. Vasquez, program manager for IEEE TechEthics.
[ IEEE TechEthics ]
Two videos this week from Lex Fridman’s AI podcast: Noam Chomsky, and Whitney Cummings.
[ AI Podcast ]
This week’s CMU RI Seminar comes from Jeff Clune at the University of Wyoming, on “Improving Robot and Deep Reinforcement Learning via Quality Diversity and Open-Ended Algorithms.”
Quality Diversity (QD) algorithms are those that seek to produce a diverse set of high-performing solutions to problems. I will describe them and a number of their positive attributes. I will then summarize our Nature paper on how they, when combined with Bayesian Optimization, produce a learning algorithm that enables robots, after being damaged, to adapt in 1-2 minutes in order to continue performing their mission, yielding state-of-the-art robot damage recovery. I will next describe our QD-based Go-Explore algorithm, which dramatically improves the ability of deep reinforcement learning algorithms to solve previously unsolvable problems wherein reward signals are sparse, meaning that intelligent exploration is required. Go-Explore solves Montezuma’s Revenge, considered by many to be a major AI research challenge. Finally, I will motivate research into open-ended algorithms, which seek to innovate endlessly, and introduce our POET algorithm, which generates its own training challenges while learning to solve them, automatically creating a curricula for robots to learn an expanding set of diverse skills. POET creates and solves challenges that are unsolvable with traditional deep reinforcement learning techniques.
[ CMU RI ] Continue reading
Posted in Human Robots
#436414 Japanese Researchers Teaching Robots to ...
When mobile manipulators eventually make it into our homes, self-repair is going to be a very important function. Hopefully, these robots will be durable enough that they won’t need to be repaired very often, but from time to time they’ll almost certainly need minor maintenance. At Humanoids 2019 in Toronto, researchers from the University of Tokyo showed how they taught a PR2 to perform simple repairs on itself by tightening its own screws. And using that skill, the robot was also able to augment itself, adding accessories like hooks to help it carry more stuff. Clever robot!
To keep things simple, the researchers provided the robot with CAD data that tells it exactly where all of its screws are.
At the moment, the robot can’t directly detect on its own whether a particular screw needs tightening, although it can tell if its physical pose doesn’t match its digital model, which suggests that something has gone wonky. It can also check its screws autonomously from time to time, or rely on a human physically pointing out that it has a screw loose, using the human’s finger location to identify which screw it is. Another challenge is that most robots, like most humans, are limited in the areas on themselves that they can comfortably reach. So to tighten up everything, they might have to find themselves a robot friend to help, just like humans help each other put on sunblock.
The actual tightening is either super easy or quite complicated, depending on the location and orientation of the screw. If the robot is lucky, it can just use its continuous wrist rotation for tightening, but if a screw is located in a tight position that requires an Allen wrench, the robot has to regrasp the tool over and over as it incrementally tightens the screw.
Image: University of Tokyo
In one experiment, the researchers taught a PR2 robot to attach a hook to one of its shoulders. The robot uses one hand to grasp the hook and another hand to grasp a screwdriver. The researchers tested the hook by hanging a tote bag on it.
The other neat trick that a robot can do once it can tighten screws on its own body is to add new bits of hardware to itself. PR2 was thoughtfully designed with mounting points on its shoulders (or maybe technically its neck) and head, and it turns out that it can reach these points with its manipulators, allowing to modify itself, as the researchers explain:
When PR2 wants to have a lot of things, the only two hands are not enough to realize that. So we let PR2 to use a bag the same as we put it on our shoulder. PR2 started attaching the hook whose pose is calculated with self CAD data with a driver on his shoulder in order to put a bag on his shoulder. PR2 finished attaching the hook, and the people put a lot of cans in a tote bag and put it on PR2’s shoulder.
“Self-Repair and Self-Extension by Tightening Screws based on Precise Calculation of Screw Pose of Self-Body with CAD Data and Graph Search with Regrasping a Driver,” by Takayuki Murooka, Kei Okada, and Masayuki Inaba from the University of Tokyo, was presented at Humanoids 2019 in Toronto, Canada. Continue reading
Posted in Human Robots
#436180 Bipedal Robot Cassie Cal Learns to ...
There’s no particular reason why knowing how to juggle would be a useful skill for a robot. Despite this, robots are frequently taught how to juggle things. Blind robots can juggle, humanoid robots can juggle, and even drones can juggle. Why? Because juggling is hard, man! You have to think about a bunch of different things at once, and also do a bunch of different things at once, which this particular human at least finds to be overly stressful. While juggling may not stress robots out, it does require carefully coordinated sensing and computing and actuation, which means that it’s as good a task as any (and a more entertaining task than most) for testing the capabilities of your system.
UC Berkeley’s Cassie Cal robot, which consists of two legs and what could be called a torso if you were feeling charitable, has just learned to juggle by bouncing a ball on what would be her head if she had one of those. The idea is that if Cassie can juggle while balancing at the same time, she’ll be better able to do other things that require dynamic multitasking, too. And if that doesn’t work out, she’ll still be able to join the circus.
Cassie’s juggling is assisted by an external motion capture system that tracks the location of the ball, but otherwise everything is autonomous. Cassie is able to juggle the ball by leaning forwards and backwards, left and right, and moving up and down. She does this while maintaining her own balance, which is the whole point of this research—successfully executing two dynamic behaviors that may sometimes be at odds with one another. The end goal here is not to make a better juggling robot, but rather to explore dynamic multitasking, a skill that robots will need in order to be successful in human environments.
This work is from the Hybrid Robotics Lab at UC Berkeley, led by Koushil Sreenath, and is being done by Katherine Poggensee, Albert Li, Daniel Sotsaikich, Bike Zhang, and Prasanth Kotaru.
For a bit more detail, we spoke with Albert Li via email.
Image: UC Berkeley
UC Berkeley’s Cassie Cal getting ready to juggle.
IEEE Spectrum: What would be involved in getting Cassie to juggle without relying on motion capture?
Albert Li: Our motivation for starting off with motion capture was to first address the control challenge of juggling on a biped without worrying about implementing the perception. We actually do have a ball detector working on a camera, which would mean we wouldn’t have to rely on the motion capture system. However, we need to mount the camera in a way that it would provide the best upwards field of view, and we also have develop a reliable estimator. The estimator is particularly important because when the ball gets close enough to the camera, we actually can’t track the ball and have to assume our dynamic models describe its motion accurately enough until it bounces back up.
What keeps Cassie from juggling indefinitely?
There are a few factors that affect how long Cassie can sustain a juggle. While in simulation the paddle exhibits homogeneous properties like its stiffness and damping, in reality every surface has anisotropic contact properties. So, there are parts of the paddle which may be better for juggling than others (and importantly, react differently than modeled). These differences in contact are also exacerbated due to how the paddle is cantilevered when mounted on Cassie. When the ball hits these areas, it leads to a larger than expected error in a juggle. Due to the small size of the paddle, the ball may then just hit the paddle’s edge and end the juggling run. Over a very long run, this is a likely occurrence. Additionally, some large juggling errors could cause Cassie’s feet to slip slightly, which ends up changing the stable standing position over time. Since this version of the controller assumes Cassie is stationary, this change in position eventually leads to poor juggles and failure.
Would Cassie be able to juggle while walking (or hovershoe-ing)?
Walking (and hovershoe-ing) while juggling is a far more challenging problem and is certainly a goal for future research. Some of these challenges include getting the paddle to precise poses to juggle the ball while also moving to avoid any destabilizing effects of stepping incorrectly. The number of juggles per step of walking could also vary and make the mathematics of the problem more challenging. The controller goal is also more involved. While the current goal of the juggling controller is to juggle the ball to a static apex position, with a walking juggling controller, we may instead want to hit the ball forwards and also walk forwards to bounce it, juggle the ball along a particular path, etc. Solving such challenges would be the main thrusts of the follow-up research.
Can you give an example of a practical task that would be made possible by using a controller like this?
Studying juggling means studying contact behavior and leveraging our models of it to achieve a known objective. Juggling could also be used to study predictable post-contact flight behavior. Consider the scenario where a robot is attempting to make a catch, but fails, letting the ball to bounce off of its hand, and then recovering the catch. This behavior could also be intentional: It is often easier to first execute a bounce to direct the target and then perform a subsequent action. For example, volleyball players could in principle directly hit a spiked ball back, but almost always bump the ball back up and then return it.
Even beyond this motivating example, the kinds of models we employ to get juggling working are more generally applicable to any task that involves contact, which could include tasks besides bouncing like sliding and rolling. For example, clearing space on a desk by pushing objects to the side may be preferable than individually manipulating each and every object on it.
You mention collaborative juggling or juggling multiple balls—is that something you’ve tried yet? Can you talk a bit more about what you’re working on next?
We haven’t yet started working on collaborative or multi-ball juggling, but that’s also a goal for future work. Juggling multiple balls statically is probably the most reasonable next goal, but presents additional challenges. For instance, you have to encode a notion of juggling urgency (if the second ball isn’t hit hard enough, you have less time to get the first ball up before you get back to the second one).
On the other hand, collaborative human-robot juggling requires a more advanced decision-making framework. To get robust multi-agent juggling, the robot will need to employ some sort of probabilistic model of the expected human behavior (are they likely to move somewhere? Are they trying to catch the ball high or low? Is it safe to hit the ball back?). In general, developing such human models is difficult since humans are fairly unpredictable and often don’t exhibit rational behavior. This will be a focus of future work.
[ Hybrid Robotics Lab ] Continue reading
Posted in Human Robots
#435822 The Internet Is Coming to the Rest of ...
People surf it. Spiders crawl it. Gophers navigate it.
Now, a leading group of cognitive biologists and computer scientists want to make the tools of the Internet accessible to the rest of the animal kingdom.
Dubbed the Interspecies Internet, the project aims to provide intelligent animals such as elephants, dolphins, magpies, and great apes with a means to communicate among each other and with people online.
And through artificial intelligence, virtual reality, and other digital technologies, researchers hope to crack the code of all the chirps, yips, growls, and whistles that underpin animal communication.
Oh, and musician Peter Gabriel is involved.
“We can use data analysis and technology tools to give non-humans a lot more choice and control,” the former Genesis frontman, dressed in his signature Nehru-style collar shirt and loose, open waistcoat, told IEEE Spectrum at the inaugural Interspecies Internet Workshop, held Monday in Cambridge, Mass. “This will be integral to changing our relationship with the natural world.”
The workshop was a long time in the making.
Eighteen years ago, Gabriel visited a primate research center in Atlanta, Georgia, where he jammed with two bonobos, a male named Kanzi and his half-sister Panbanisha. It was the first time either bonobo had sat at a piano before, and both displayed an exquisite sense of musical timing and melody.
Gabriel seemed to be speaking to the great apes through his synthesizer. It was a shock to the man who once sang “Shock the Monkey.”
“It blew me away,” he says.
Add in the bonobos’ ability to communicate by pointing to abstract symbols, Gabriel notes, and “you’d have to be deaf, dumb, and very blind not to notice language being used.”
Gabriel eventually teamed up with Internet protocol co-inventor Vint Cerf, cognitive psychologist Diana Reiss, and IoT pioneer Neil Gershenfeld to propose building an Interspecies Internet. Presented in a 2013 TED Talk as an “idea in progress,” the concept proved to be ahead of the technology.
“It wasn’t ready,” says Gershenfeld, director of MIT’s Center for Bits and Atoms. “It needed to incubate.”
So, for the past six years, the architects of the Dolittlesque initiative embarked on two small pilot projects, one for dolphins and one for chimpanzees.
At her Hunter College lab in New York City, Reiss developed what she calls the D-Pad—a touchpad for dolphins.
Reiss had been trying for years to create an underwater touchscreen with which to probe the cognition and communication skills of bottlenose dolphins. But “it was a nightmare coming up with something that was dolphin-safe and would work,” she says.
Her first attempt emitted too much heat. A Wii-like system of gesture recognition proved too difficult to install in the dolphin tanks.
Eventually, she joined forces with Rockefeller University biophysicist Marcelo Magnasco and invented an optical detection system in which images and infrared sensors are projected through an underwater viewing window onto a glass panel, allowing the dolphins to play specially designed apps, including one dubbed Whack-a-Fish.
Meanwhile, in the United Kingdom, Gabriel worked with Alison Cronin, director of the ape rescue center Monkey World, to test the feasibility of using FaceTime with chimpanzees.
The chimps engaged with the technology, Cronin reported at this week’s workshop. However, our hominid cousins proved as adept at videotelephonic discourse as my three-year-old son is at video chatting with his grandparents—which is to say, there was a lot of pass-the-banana-through-the-screen and other silly games, and not much meaningful conversation.
“We can use data analysis and technology tools to give non-humans a lot more choice and control.”
—Peter Gabriel
The buggy, rudimentary attempt at interspecies online communication—what Cronin calls her “Max Headroom experiment”—shows that building the Interspecies Internet will not be as simple as giving out Skype-enabled tablets to smart animals.
“There are all sorts of problems with creating a human-centered experience for another animal,” says Gabriel Miller, director of research and development at the San Diego Zoo.
Miller has been working on animal-focused sensory tools such as an “Elephone” (for elephants) and a “Joybranch” (for birds), but it’s not easy to design efficient interactive systems for other creatures—and for the Interspecies Internet to be successful, Miller points out, “that will be super-foundational.”
Researchers are making progress on natural language processing of animal tongues. Through a non-profit organization called the Earth Species Project, former Firefox designer Aza Raskin and early Twitter engineer Britt Selvitelle are applying deep learning algorithms developed for unsupervised machine translation of human languages to fashion a Rosetta Stone–like tool capable of interpreting the vocalizations of whales, primates, and other animals.
Inspired by the scientists who first documented the complex sonic arrangements of humpback whales in the 1960s—a discovery that ushered in the modern marine conservation movement—Selvitelle hopes that an AI-powered animal translator can have a similar effect on environmentalism today.
“A lot of shifts happen when someone who doesn’t have a voice gains a voice,” he says.
A challenge with this sort of AI software remains verification and validation. Normally, machine-learning algorithms are benchmarked against a human expert, but who is to say if a cybernetic translation of a sperm whale’s clicks is accurate or not?
One could back-translate an English expression into sperm whale-ese and then into English again. But with the great apes, there might be a better option.
According to primatologist Sue Savage-Rumbaugh, expertly trained bonobos could serve as bilingual interpreters, translating the argot of apes into the parlance of people, and vice versa.
Not just any trained ape will do, though. They have to grow up in a mixed Pan/Homo environment, as Kanzi and Panbanisha were.
“If I can have a chat with a cow, maybe I can have more compassion for it.”
—Jeremy Coller
Those bonobos were raised effectively from birth both by Savage-Rumbaugh, who taught the animals to understand spoken English and to communicate via hundreds of different pictographic “lexigrams,” and a bonobo mother named Matata that had lived for six years in the Congolese rainforests before her capture.
Unlike all other research primates—which are brought into captivity as infants, reared by human caretakers, and have limited exposure to their natural cultures or languages—those apes thus grew up fluent in both bonobo and human.
Panbanisha died in 2012, but Kanzi, aged 38, is still going strong, living at an ape sanctuary in Des Moines, Iowa. Researchers continue to study his cognitive abilities—Francine Dolins, a primatologist at the University of Michigan-Dearborn, is running one study in which Kanzi and other apes hunt rabbits and forage for fruit through avatars on a touchscreen. Kanzi could, in theory, be recruited to check the accuracy of any Google Translate–like app for bonobo hoots, barks, grunts, and cries.
Alternatively, Kanzi could simply provide Internet-based interpreting services for our two species. He’s already proficient at video chatting with humans, notes Emily Walco, a PhD student at Harvard University who has personally Skyped with Kanzi. “He was super into it,” Walco says.
And if wild bonobos in Central Africa can be coaxed to gather around a computer screen, Savage-Rumbaugh is confident Kanzi could communicate with them that way. “It can all be put together,” she says. “We can have an Interspecies Internet.”
“Both the technology and the knowledge had to advance,” Savage-Rumbaugh notes. However, now, “the techniques that we learned could really be extended to a cow or a pig.”
That’s music to the ears of Jeremy Coller, a private equity specialist whose foundation partially funded the Interspecies Internet Workshop. Coller is passionate about animal welfare and has devoted much of his philanthropic efforts toward the goal of ending factory farming.
At the workshop, his foundation announced the creation of the Coller Doolittle Prize, a US $100,000 award to help fund further research related to the Interspecies Internet. (A working group also formed to synthesize plans for the emerging field, to facilitate future event planning, and to guide testing of shared technology platforms.)
Why would a multi-millionaire with no background in digital communication systems or cognitive psychology research want to back the initiative? For Coller, the motivation boils to interspecies empathy.
“If I can have a chat with a cow,” he says, “maybe I can have more compassion for it.”
An abridged version of this post appears in the September 2019 print issue as “Elephants, Dolphins, and Chimps Need the Internet, Too.” 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