In case you have been hiding under a rock, Artificial Intelligence (AI) has been in the news a lot lately. It has been used to train robots to work in complex settings and to reduce the amount data needed to generate ocular images in a clinical setting.
Training Robots
Two studies done at the University of Washington used AI systems to train robots for real life settings using either video or photos to create simulations. This will lower the cost of training robots.
In the first study, someone scans an area with a smartphone to record its geometry. A system, called RialTo, creates a digital twin simulation of the space. This system allows a user to enter how things, like a drawer, function. A robot can virtually repeat motions in simulation to learn how to do them successfully. In the second study, researchers built a system called URDFormer. This system takes images from the Internet of real environments and creates physically realistic simulation environments where robots can train.
With both studies, researchers were working to allow systems to inexpensively go from real world to simulation, so that the systems can train the robot and help it function better in a physical space. This is good for safety, since no one wants a poorly trained robot breaking things or hurting people. This also lowers the barrier for access, especially if you can have a robot work on your house by just scanning the area in question with your phone.
While robots have been working successfully in assembly lines for many years, teaching them to interact in less structured surroundings has been a problem. These two systems approach the problem of having a robot work in an area, like someone’s home in different ways.
RialTo, which was created with a team from the Massachusetts Institute of Technology (MIT), has a person move through a space and take video of its geometry and moving parts, such as opening cabinets or the refrigerator door. This system uses existing AI models and a human does some work through a user interface to show how things move in order to create a simulated version of the space. A virtual robot trains itself through trial and error, which is known as reinforcement learning. By doing this process in the simulation, it improves at that task and can transfer that learning to a physical environment. Robots that are trained with the RialTo system are almost as accurate as a robot trained in a real space, like a kitchen.
The other system, URDFormer, is less focused on accuracy in a single kitchen. It is focused on quickly and cheaply forms hundreds of generic kitchen simulations. This system scans images from the Internet and pairs them with existing models of how things like kitchen drawers or cabinets work. Next, it predicts a simulation from the real-world image. This allows researchers to train robots, quickly and inexpensively in a wide range of settings. The downside is that the simulations are less accurate than those made by the RialTo system.
The hope is to have both systems complement each other. The URDFormer is useful for pre-training on many scenarios. The RialTo system is good if there is a pre-trained robot that will be used in someone’s home and the goal is to have it be 95 percent successful.
AI in the Clinic
Using AI to train robots is one thing. What about using AI in medical settings, such as optometry and ophthalmology offices? Researchers at the National Institutes of Health (NIH) found that by combining AI with imaging, they were able to reduce the data required for imaging technology to view individual cells in the eye. This work is helping to make this kind of technology available for use in eye clinics.
Many of the causes of blindness involve the retina. Recently there have been advances in the imaging of the eye, known as adaptive optics. This has made it possible to obtain detailed 3D images of retinal cells. When it is combined with technologies, like optical coherence tomography (OCT), it can acquire assessment of the retina at the cellular level. This high-level imaging can help eye doctors to determine if particular treatment is working.
The trouble is the current use of adaptive optics imaging with OCT are impractical for day-to-day use. It requires a patient to hold their eye very still while hundreds of images of the retina are taken. This is difficult for both the patient and the clinic, since a large amount of imaging data is generated. When using adaptive optics, it’s a tradeoff between taking fewer measurement for efficiency’s sake and getting a good-resolution image.
Enter the AI system known as residual-inresidual transformer generative adversarial network (RRTGAN). This restores pixel resolution from fewer measurements. It requires one-fourth of the data that adaptive optics require, and it enables a quick, cellular-scale 3D image of the eye in a clinical setting.
“Getting the most advanced ophthalmic imaging technologies into the hands of healthcare providers will vastly improve the ability to detect retinal diseases earlier, and guide treatments to prevent vision loss,” said Johnny Tam, Ph.D., investigator at NIH’s National Eye Institute (NEI) and senior author of the study report, which published in the Nature journal, NPJ Artificial Intelligence.
While there is concern about AI use, and rightly so, it can be used to help train robots and to get high-quality images of the eye more efficiently than current methods. Like other technology that is being utilized, AI is a tool that needs human oversight in order to use it property. When used properly, AI can do a great deal of good.
Sources:
https://www.washington.edu/news/2024/08/07/ai-robots-reinforcement-learning-training-simulation/