Connect with us

TOP SCEINCE

A technique for more effective multipurpose robots

Published

on

A technique for more effective multipurpose robots


Let’s say you want to train a robot so it understands how to use tools and can then quickly learn to make repairs around your house with a hammer, wrench, and screwdriver. To do that, you would need an enormous amount of data demonstrating tool use.

Existing robotic datasets vary widely in modality — some include color images while others are composed of tactile imprints, for instance. Data could also be collected in different domains, like simulation or human demos. And each dataset may capture a unique task and environment.

It is difficult to efficiently incorporate data from so many sources in one machine-learning model, so many methods use just one type of data to train a robot. But robots trained this way, with a relatively small amount of task-specific data, are often unable to perform new tasks in unfamiliar environments.

In an effort to train better multipurpose robots, MIT researchers developed a technique to combine multiple sources of data across domains, modalities, and tasks using a type of generative AI known as diffusion models.

They train a separate diffusion model to learn a strategy, or policy, for completing one task using one specific dataset. Then they combine the policies learned by the diffusion models into a general policy that enables a robot to perform multiple tasks in various settings.

In simulations and real-world experiments, this training approach enabled a robot to perform multiple tool-use tasks and adapt to new tasks it did not see during training. The method, known as Policy Composition (PoCo), led to a 20 percent improvement in task performance when compared to baseline techniques.

“Addressing heterogeneity in robotic datasets is like a chicken-egg problem. If we want to use a lot of data to train general robot policies, then we first need deployable robots to get all this data. I think that leveraging all the heterogeneous data available, similar to what researchers have done with ChatGPT, is an important step for the robotics field,” says Lirui Wang, an electrical engineering and computer science (EECS) graduate student and lead author of a paper on PoCo.

Wang’s coauthors include Jialiang Zhao, a mechanical engineering graduate student; Yilun Du, an EECS graduate student; Edward Adelson, the John and Dorothy Wilson Professor of Vision Science in the Department of Brain and Cognitive Sciences and a member of the Computer Science and Artificial Intelligence Laboratory (CSAIL); and senior author Russ Tedrake, the Toyota Professor of EECS, Aeronautics and Astronautics, and Mechanical Engineering, and a member of CSAIL. The research will be presented at the Robotics: Science and Systems Conference.

Combining disparate datasets

A robotic policy is a machine-learning model that takes inputs and uses them to perform an action. One way to think about a policy is as a strategy. In the case of a robotic arm, that strategy might be a trajectory, or a series of poses that move the arm so it picks up a hammer and uses it to pound a nail.

Datasets used to learn robotic policies are typically small and focused on one particular task and environment, like packing items into boxes in a warehouse.

“Every single robotic warehouse is generating terabytes of data, but it only belongs to that specific robot installation working on those packages. It is not ideal if you want to use all of these data to train a general machine,” Wang says.

The MIT researchers developed a technique that can take a series of smaller datasets, like those gathered from many robotic warehouses, learn separate policies from each one, and combine the policies in a way that enables a robot to generalize to many tasks.

They represent each policy using a type of generative AI model known as a diffusion model. Diffusion models, often used for image generation, learn to create new data samples that resemble samples in a training dataset by iteratively refining their output.

But rather than teaching a diffusion model to generate images, the researchers teach it to generate a trajectory for a robot. They do this by adding noise to the trajectories in a training dataset. The diffusion model gradually removes the noise and refines its output into a trajectory.

This technique, known as Diffusion Policy, was previously introduced by researchers at MIT, Columbia University, and the Toyota Research Institute. PoCo builds off this Diffusion Policy work.

The team trains each diffusion model with a different type of dataset, such as one with human video demonstrations and another gleaned from teleoperation of a robotic arm.

Then the researchers perform a weighted combination of the individual policies learned by all the diffusion models, iteratively refining the output so the combined policy satisfies the objectives of each individual policy.

Greater than the sum of its parts

“One of the benefits of this approach is that we can combine policies to get the best of both worlds. For instance, a policy trained on real-world data might be able to achieve more dexterity, while a policy trained on simulation might be able to achieve more generalization,” Wang says.

Because the policies are trained separately, one could mix and match diffusion policies to achieve better results for a certain task. A user could also add data in a new modality or domain by training an additional Diffusion Policy with that dataset, rather than starting the entire process from scratch.

The researchers tested PoCo in simulation and on real robotic arms that performed a variety of tools tasks, such as using a hammer to pound a nail and flipping an object with a spatula. PoCo led to a 20 percent improvement in task performance compared to baseline methods.

“The striking thing was that when we finished tuning and visualized it, we can clearly see that the composed trajectory looks much better than either one of them individually,” Wang says.

In the future, the researchers want to apply this technique to long-horizon tasks where a robot would pick up one tool, use it, then switch to another tool. They also want to incorporate larger robotics datasets to improve performance.

“We will need all three kinds of data to succeed for robotics: internet data, simulation data, and real robot data. How to combine them effectively will be the million-dollar question. PoCo is a solid step on the right track,” says Jim Fan, senior research scientist at NVIDIA and leader of the AI Agents Initiative, who was not involved with this work.

This research is funded, in part, by Amazon, the Singapore Defense Science and Technology Agency, the U.S. National Science Foundation, and the Toyota Research Institute.



Source link

Continue Reading
Click to comment

Leave a Reply

TOP SCEINCE

Prying open the AI black box

Published

on

By

Prying open the AI black box


Artificial intelligence continues to squirm its way into many aspects of our lives. But what about biology, the study of life itself? AI can sift through hundreds of thousands of genome data points to identify potential new therapeutic targets. While these genomic insights may appear helpful, scientists aren’t sure how today’s AI models come to their conclusions in the first place. Now, a new system named SQUID arrives on the scene armed to pry open AI’s black box of murky internal logic.

SQUID, short for Surrogate Quantitative Interpretability for Deepnets, is a computational tool created by Cold Spring Harbor Laboratory (CSHL) scientists. It’s designed to help interpret how AI models analyze the genome. Compared with other analysis tools, SQUID is more consistent, reduces background noise, and can lead to more accurate predictions about the effects of genetic mutations.

How does it work so much better? The key, CSHL Assistant Professor Peter Koo says, lies in SQUID’s specialized training.

“The tools that people use to try to understand these models have been largely coming from other fields like computer vision or natural language processing. While they can be useful, they’re not optimal for genomics. What we did with SQUID was leverage decades of quantitative genetics knowledge to help us understand what these deep neural networks are learning,” explains Koo.

SQUID works by first generating a library of over 100,000 variant DNA sequences. It then analyzes the library of mutations and their effects using a program called MAVE-NN (Multiplex Assays of Variant Effects Neural Network). This tool allows scientists to perform thousands of virtual experiments simultaneously. In effect, they can “fish out” the algorithms behind a given AI’s most accurate predictions. Their computational “catch” could set the stage for experiments that are more grounded in reality.

“In silico [virtual] experiments are no replacement for actual laboratory experiments. Nevertheless, they can be very informative. They can help scientists form hypotheses for how a particular region of the genome works or how a mutation might have a clinically relevant effect,” explains CSHL Associate Professor Justin Kinney, a co-author of the study.

There are tons of AI models in the sea. More enter the waters each day. Koo, Kinney, and colleagues hope that SQUID will help scientists grab hold of those that best meet their specialized needs.

Though mapped, the human genome remains an incredibly challenging terrain. SQUID could help biologists navigate the field more effectively, bringing them closer to their findings’ true medical implications.



Source link

Continue Reading

TOP SCEINCE

Iron meteorites hint that our infant solar system was more doughnut than dartboard

Published

on

By

Iron meteorites hint that our infant solar system was more doughnut than dartboard


Four and a half billion years ago, our solar system was a cloud of gas and dust swirling around the sun, until gas began to condense and accrete along with dust to form asteroids and planets. What did this cosmic nursery, known as a protoplanetary disk, look like, and how was it structured? Astronomers can use telescopes to “see” protoplanetary disks far away from our much more mature solar system, but it is impossible to observe what ours might have looked like in its infancy — only an alien billions of light years away would be able to see it as it once was.

Fortunately, space has dropped a few clues — fragments of objects that formed early in solar system history and plunged through Earth’s atmosphere, called meteorites. The composition of meteorites tells stories of the solar system’s birth, but these stories often raise more questions than answers.

In a paper published in Proceedings of the National Academy of Sciences, a team of planetary scientists from UCLA and Johns Hopkins University Applied Physics Laboratory reports that refractory metals, which condense at high temperatures, such as iridium and platinum, were more abundant in meteorites formed in the outer disk, which was cold and far away from the sun. These metals should have formed close to the sun, where the temperature was much higher. Was there a pathway that moved these metals from the inner disk to the outer?

Most meteorites formed within the first few million years of solar system history. Some meteorites, called chondrites, are unmelted conglomerations of grains and dust left over from planet formation. Other meteorites experienced enough heat to melt while their parent asteroids were forming. When these asteroids melted, the silicate part and the metallic part separated due to their difference in density, similar to how water and oil don’t mix.

Today, most asteroids are located in a thick belt between Mars and Jupiter. Scientists think that Jupiter’s gravity disrupted the course of these asteroids, causing many of them to smash into each other and break apart. When pieces of these asteroids fall to Earth and are recovered, they are called meteorites.

Iron meteorites are from the metallic cores of the earliest asteroids, older than any other rocks or celestial objects in our solar system. The irons contain molybdenum isotopes that point toward many different locations across the protoplanetary disk in which these meteorites formed. That allows scientists to learn what the chemical composition of the disk was like in its infancy.

Previous research using the Atacama Large Millimeter/submillimeter Array in Chile has found many disks around other stars that resemble concentric rings, like a dartboard. The rings of these planetary disks, such as HL Tau, are separated by physical gaps, so this kind of disk could not provide a route to transport these refractory metals from the inner disk to the outer.

The new paper holds that our solar disk likely didn’t have a ring structure at the very beginning. Instead, our planetary disk looked more like a doughnut, and asteroids with metal grains rich in iridium and platinum metals migrated to the outer disk as it rapidly expanded.

But that confronted the researchers with another puzzle. After the disk expansion, gravity should have pulled these metals back into the sun. But that did not happen.

“Once Jupiter formed, it very likely opened a physical gap that trapped the iridium and platinum metals in the outer disk and prevented them from falling into the sun,” said first author Bidong Zhang, a UCLA planetary scientist. “These metals were later incorporated into asteroids that formed in the outer disk. This explains why meteorites formed in the outer disk — carbonaceous chondrites and carbonaceous-type iron meteorites — have much higher iridium and platinum contents than their inner-disk peers.”

Zhang and his collaborators previously used iron meteorites to reconstruct how water was distributed in the protoplanetary disk.

“Iron meteorites are hidden gems. The more we learn about iron meteorites, the more they unravel the mystery of our solar system’s birth,” Zhang said.



Source link

Continue Reading

TOP SCEINCE

Supermassive black hole appears to grow like a baby star

Published

on

By

Supermassive black hole appears to grow like a baby star


Supermassive black holes pose unanswered questions for astronomers around the world, not least “How do they grow so big?” Now, an international team of astronomers, including researchers from Chalmers University of Technology in Sweden, has discovered a powerful rotating, magnetic wind that they believe is helping a galaxy’s central supermassive black hole to grow. The swirling wind, revealed with the help of the ALMA telescope in nearby galaxy ESO320-G030, suggests that similar processes are involved both in black hole growth and the birth of stars.

Most galaxies, including our own Milky Way have a supermassive black hole at their centre. How these mind-bogglingly massive objects grow to weigh as much as millions or billions of stars is a long-standing question for astronomers.

In search of clues to this mystery, a team of scientists led by Mark Gorski (Northwestern University and Chalmers) and Susanne Aalto (Chalmers) chose to study the relatively nearby galaxy ESO320-G030, only 120 million light years distant. It’s a very active galaxy, forming stars ten times as fast as in our own galaxy.

“Since this galaxy is very luminous in the infrared, telescopes can resolve striking details in its centre. We wanted to measure light from molecules carried by winds from the galaxy’s core, hoping to trace how the winds are launched by a growing, or soon to be growing, supermassive black hole. By using ALMA, we were able to study light from behind thick layers of dust and gas,” says Susanne Aalto, Professor of Radio Astronomy at Chalmers University of Technology.

To zero in on dense gas from as close as possible to the central black hole, the scientists studied light from molecules of hydrogen cyanide (HCN). Thanks to ALMA’s ability to image fine details and trace movements in the gas — using the Doppler effect — they discovered patterns that suggest the presence of a magnetised, rotating wind.

While other winds and jets in the centre of galaxies push material away from the supermassive black hole, the newly discovered wind adds another process, that can instead feed the black hole and help it grow.

“We can see how the winds form a spiralling structure, billowing out from the galaxy’s centre. When we measured the rotation, mass, and velocity of the material flowing outwards, we were surprised to find that we could rule out many explanations for the power of the wind, star formation for example. Instead, the flow outwards may be powered by the inflow of gas and seems to be held together by magnetic fields,” says Susanne Aalto.

The scientists think that the rotating magnetic wind helps the black hole to grow.

Material travels around the black hole before it can fall in — like water around a drain. Matter that approaches the black hole collects in a chaotic, spinning disk. There, magnetic fields develop and get stronger. The magnetic fields help lift matter away from the galaxy, creating the spiralling wind. Losing matter to this wind also slows the spinning disk — that means that matter can flow more easily into the black hole, turning a trickle into a stream.

For Mark Gorski, the way this happens is strikingly reminiscent of a much smaller-scale environment in space: the swirls of gas and dust that lead up to the birth of new stars and planets.

“It is well-established that stars in the first stages of their evolution grow with the help of rotating winds — accelerated by magnetic fields, just like the wind in this galaxy. Our observations show that supermassive black holes and tiny stars can grow by similar processes, but on very different scales,” says Mark Gorski.

Could this discovery be a clue to solving the mystery of how supermassive black holes grow? In the future, Mark Gorski, Susanne Aalto and their colleagues want to study other galaxies which may harbour hidden spiralling outflows in their centres.

“Far from all questions about this process are answered. In our observations we see clear evidence of a rotating wind that helps regulate the growth of the galaxy’s central black hole. Now that we know what to look for, the next step is to find out how common a phenomenon this is. And if this is a stage which all galaxies with supermassive black holes go through, what happens to them next?,” asks Mark Gorski.



Source link

Continue Reading

Trending

Copyright © 2017 Zox News Theme. Theme by MVP Themes, powered by WordPress.