AI Advancement Roundup: 4 Cool Developments

For people it’s relatively easy to move items from a pile to find something underneath. Without even thinking about it, we consider the shape of the object and the likelihood of where it might be. We then make a series of choices as we search, preventing things from crushing each other or falling to the ground.

Could a robot find a lost item using the same logic? An MIT research team attempted to find out. Their solution, a system they named FuseBot, was taught to use visual cues, RFID tag data, and probabilistic reasoning to analyse the amount of space an item was likely to take up and strategically move objects to reach it. 180 trials were run using a variety of general household items and the targeted objects were found 95% of the time. FuseBot was even taught to find items missing the RFID tag in as few moves as possible by scanning the pile as a whole and using the same logic while searching where there was not a tag.

It’s unlikely that this technology is necessary to help us find our keys in the morning, so how could FuseBot be used? The speed and efficiency of this system would be especially useful in sorting and processing returns (both with and without RFID tags) in an ecommerce warehouse as the reasoning algorithm of FuseBot finds hidden objects with more accuracy in half the time of existing robotic systems.

The ocean is unpredictable and can be quite dangerous, but MIT engineers were undaunted in their attempt to find a pattern in the chaos. Starting with standard equations of simple wave behaviour, the scientists used AI and wave-tank data to more accurately predict the complexity of breaking waves in the deep ocean.

The machine-learning algorithm was trained to compare the movement of waves created in a tank with calculated potential waves, determine the difference, and then apply this data to predict new, complex situations. The trained model was able to make more accurate forecasts of change in movement, speed, and size of waves than the untrained model.

The team’s updated model could be used for more in-depth analysis of wave behaviour as it reacts with offshore structures. By simulating how waves break using these new methods, engineers can better test the resilience of various designs and materials to build safer offshore structures more efficiently. This model can also help with climate research and allow scientists to more easily predict how much carbon dioxide the ocean can absorb.

MIT Media Lab has found a way to make more intelligent textiles. Smart fabrics have been manufactured in the past using a combination of yarn and functional fibre, but the pliability of the material generates noise that affects the sensitivity of pressure sensors that analyse the data, leading to less-than-optimal results. By incorporating thermoplastic yarns that harden with heat, engineers developed 3DKnITS, a new material that can be formed into more precise shapes while providing more accurate data through the use of a newly developed AI model.

The scientists created a smart, form-fitting shoe and mat out of the new material and used the new model to interpret the data from the pressure sensors in real time. The system was able to predict motion and yoga poses with 99% accuracy.

While the large-scale fabrication process of thermoforming yarns into material already exists, the machine-learning model currently needs to be calibrated to each individual. As refinements are made to the process, this technology could be used to help orthopaedic and sports medicine doctors treat existing injuries and identify existing force distribution issues to prevent future injuries.

Manipulating dough can be difficult for humans, with a long sequence of steps that require choices along the way. Should you use your hands or a roller? Does the dough need to be turned to get the right shape? Is it thin enough? There are no specific, clear steps to follow as the process requires many decisions and a bit of trial and error along the way.  

Robots can’t solve problems through trial and error, but researchers from MIT, Carnegie Mellon University, and the University of California San Diego have come up with a two-stage learning process to attempt to replicate this human concept. First an algorithm was developed to break the overall task down into smaller pieces and then a machine-learning model was taught to make decisions for when to use each skill in the process through simulations.

While pizza chef jobs are probably stills safe from robot takeover, this teacher/student machine learning model called DiffSkill could be used to train machines to manipulate a variety of deformable objects such as cloth during the manufacturing process or in the form of a caregiver robot that helps feed or dress the elderly or people with impairments.

At first glance, researching the solution to a simple problem might not seem important, but the potential exists with each AI advancement to make significant changes in our daily lives, health, and well-being.