The CSAIL team has already used the system in hospitals and assisted living facilities to monitor people for issues including Parkinson’s, dementia and COVID-19. The researchers have improved the system, which uses deep machine learning. It can identify activities, such as sleeping, reading, cooking and watching TV, and items like laptops. RF-Diary is accurate in classifying more than 30 household activities over 90 percent of the time, according to the researchers.

It uses a floor map of a subject’s living space to determine what actions they undertake in different parts of the home, and what objects they use to do so. To set up the system, the person who RF-Diary is monitoring has to carry out several actions. The system will also observe them walking around their living space to make sure it doesn’t monitor any locations the person doesn’t have access to, since radio signals can travel through walls.

Beyond protecting privacy, testing showed that the system is more effective at tracking someone’s activities in dark and “occluded” settings than video-based systems. Radio signals don’t need light, after all. The researchers plan to adapt the system for homes and hospitals, with the aim of selling it commercially.

The researchers trained the system on multiple tasks, including looking at chest X-rays to diagnose conditions like a collapsed lung. When asked to diagnose cardiomegaly (an enlarged heart), the human-AI hybrid model performed eight percent better than either the AI or medical professionals could on their own.

“There are many obstacles that understandably prohibit full automation in clinical settings, including issues of trust and accountability,” says David Sontag, lead author of a paper that the CSAIL team presented at the International Conference on Machine Learning. “We hope that our method will inspire machine learning practitioners to get more creative in integrating real-time human expertise into their algorithms.”

Next, the researchers will test a system that works with and defers to several experts at once. For instance, the AI might collaborate with different radiologists who are more experienced with different patient populations. 

The team also believes their system could have implications for content moderation because it’s able to detect offensive text and images. As social media companies struggle to remove misinformation and hate, a tool like this could help alleviate some of the burden on content moderators without resorting to full automation.

The researchers believe the approach could be used to autonomously disinfect other environments, like factories, restaurants, supermarkets and schools. The system is capable of mapping a given space, and it can navigate between waypoints and other specified areas.

“As we drive the robot around the food bank, we are also researching new control policies that will allow the robot to adapt to changes in the environment and ensure all areas receive the proper estimated dosage,” Alyssa Pierson, CSAIL research scientist and a technical lead on this project, said in a statement. “We are focused on remote operation to minimize human supervision, and therefore, the additional risk of spreading Covid-19, while running our system.” 

UVC light is sometimes used to sterilize patient rooms and other medical settings. The MTA is currently testing UV light boxes as a way to disinfect subways and buses. We’ve also seen UV light used to clean humidifiers and airplane bathrooms.

“We are excited to see the UVC disinfecting robot support our community in this time of need,” said CSAIL director and project lead Daniela Rus. “The insights we received from the work at GBFB has highlighted several algorithmic challenges. We plan to tackle these in order to extend the scope of autonomous UV disinfection in complex spaces, including dorms, schools, airplanes, and grocery stores.”

You can’t go back in time to see how Monet or Van Gogh made their masterpieces, but AI might give you the next best thing. MIT CSAIL researchers have created a machine learning system, Timecraft, that can deduce how a painting was produced and recreate the likely brushstrokes, even for famous artists. The design was first trained on 200 timelapse videos of digital and watercolor paintings, after which the scientists produced a convolutional neural network to ‘deconstruct’ artwork based on what it had learned.

The results aren’t perfect, but they’re more effective than you might think. Timecraft was better than existing benchmark tests over 90 percent of the time. And when used to recreate paintings that already have timelapse videos, it fooled almost half of the people participating in an online survey.

CSAIL’s tech isn’t ready for real-world use. A Parrot Bebop 2 drone responded to 82 percent of over 1,500 gestures — promising, but not what you’d depend on in a vital situation. The scientists intend to refine the technology, though, including the option of custom or more continuous gestures. They’ll ideally learn from the commands to better understand input or learn to navigate on their own.

If the technology does escape the lab, though, it could make robot control more accessible to people who’d otherwise be intimidated. It could also be helpful for remote exploration, personal robots and other tasks where you may want the more organic control of a human for tricky situations.

The CSAIL team has already put Emerald to use at an assisted living facility, where they used it to remotely monitor a COVID-19 patient. As the patient recovered, the system detected that her breathing rate decreased from 23 to 18 breaths per minute, her sleep improved and she was walking more quickly around her apartment.

“Given how Emerald can generate important health data without any patient contact, it could minimize the risk that doctors and nurses will catch the disease from their patients,” says Dr. Ipsit Vahia, an assistant professor of psychiatry at Harvard Medical School. That could be especially helpful in places like skilled nursing and assisted living facilities, where so many patients are at a high risk for contracting COVID-19.

BREAKING: this MIT wireless device has just been used by Boston doctors to monitor COVID-19 patients from a distance, to reduce the risk of contagion.
VIDEO: https://t.co/9g1kgNJ54u
(w/@McLeanHospital)#COVID19 #remotehealth #wireless pic.twitter.com/qNvSWBYgYu

— MIT CSAIL (@MIT_CSAIL) April 14, 2020

As the number of COVID-19 cases spike, Emerald could allow less severe patients to stay at home but remain under the supervision of healthcare providers. In the future, Emerald could be used to monitor other conditions, like anxiety, insomnia and sleep apnea. And along with telehealth, it could spur the shift toward tech-driven remote care.

Most autonomous vehicles use LIDAR sensors and/or cameras to figure out where they are on the road, but cameras can be thrown off by lighting conditions or snow-covered signs and lane markings, and LIDAR often becomes less accurate in inclement weather. GPR, on the other hand, sends electromagnetic pulses into the ground to measure the specific combination of soil, rocks and roots. That data is turned into a map for self-driving vehicles.

The system, which uses a type of GPR called Localizing Ground Penetrating Radar developed at the MIT Lincoln Laboratory, offers a few benefits. For starters, it doesn’t matter if the road is snow-covered or if visibility is blocked by fog. And conditions under the road tend to change less often than features like lane striping and signage.

“If you or I grabbed a shovel and dug it into the ground, all we’re going to see is a bunch of dirt,” says CSAIL PhD student Teddy Ort. “But LGPR can quantify the specific elements there and compare that to the map it’s already created, so that it knows exactly where it is, without needing cameras or lasers.”

So far, the CSAIL team has only tested the system at low speeds on a closed country road, but the researchers believe it could be easily extended to highways and other high-speed areas. They admit that the system doesn’t work as well in rainy conditions, when water has seeped into the ground below the road, and that it is far from road-ready. It would also have to be used in combination with other technology, as it wouldn’t detect hazards on the road.

A paper on the project will be published in the IEEE Robotics and Automation Letters journal later this month. The team plans to continue refining the hardware, so that it is less bulky — it’s currently six feet wide — and improving LGPR mapping techniques.

The array would be relatively inexpensive at just a few cents per antenna, and it would consume little power compared to a conventional system. You wouldn’t need amplifiers or other hardware that typically drains batteries, after all.

Tere’s no mention of how soon you could expect RFocus in use. The team would not only have to refine the design, but find a way to produce it at scale. The uses are already clear, at least. At a minimum, this could provide stronger, longer-ranged connections for everything from WiFi to notoriously finicky high-band 5G. However, this may be most useful for Internet of Things devices that are either too small to have a wireless link or need some additional bulk to maintain reliable signals. You could see wireless data in more devices, or more elegant versions of the gadgets you already have.

In their experiment, the team filmed a pile of clutter. Off screen, someone created shadows by moving blocks and other objects. Their algorithm predicted the light transport, or the way light is expected to move in a scene, and compared that to the shadows. It then used that info to reconstruct the off-screen video.

While the results of the work are still blurry and unrefined — the reconstructed videos show color and motion but not detail — the system could one day help self-driving cars detect what’s happening around corners or improve search-and-rescue missions in obstructed areas.

This isn’t the first time MIT has attempted to see around corners. This method improves upon that work because it doesn’t require laser-powered cameras, and it can recreate an off-screen image using any video scene, not just video of changes in lighting on the floor. Next, the CSAIL team plans to improve the resolution of their reconstructed video and to test the video in uncontrolled environments.

New research from MIT’s Computer Science and Artificial Intelligence Laboratory (CSAIL) examines the problem of how a self-driving car can predict the behavior of other drivers on the road. This prediction requires a degree of social awareness which is difficult for machines, so the researchers took tools from social psychology to help the system classify driving behaviors into either selfish and selfless.

The system observed human driving behaviors and was then able to better predict the movements of other cars when it came to merging lanes or making unprotected left turns, with 25 percent greater accuracy than previously.

This kind of insight into human behavior is important for safety when autonomous and human drivers are sharing the road. An Uber self-driving car which struck and killed a pedestrian last year, for example, didn’t have the ability to recognize jaywalkers.

“Working with and around humans means figuring out their intentions to better understand their behavior,” said graduate student Wilko Schwarting, lead author on the new paper. “People’s tendencies to be collaborative or competitive often spills over into how they behave as drivers. In this paper we sought to understand if this was something we could actually quantify.”

The research needs to be expanded before it can be implemented on real roads. The next step is for the team to apply their model to other road users like pedestrians, cyclists and other robotic systems.