New AI Research Could Enable Ocean Monitoring and Exploration

Further exploration of the unknown worlds of the deep sea may soon be possible, along with low-energy monitoring. Engineers at Caltech, ETH Zurich, and Harvard are developing an AI that will allow autonomous drones to navigate using the ocean’s currents rather than struggling to move through them.

According to one of the authors of a recent paper published in Nature Communications, John O. Dabiri, communications with robots exploring at depths of 20,000 feet is near impossible. They are difficult to control via joystick, and researchers are unable to feed them data about local ocean currents. So it becomes necessary for ocean-borne drones to be able to make their own decisions and move by themselves.

The algorithm uses reinforcement learning (RL) networks, which don’t rely on static data for training but train as fast as they collect experience. This also means that not as much computing power is needed — the team installed and ran the software on a Teensy microcontroller. The hope was for robots that could learn the types of navigation strategies already employed by animals, such as the way a fish swims or an eagle rides thermals in the air. However, by exploiting repeated trials, the RL network allowed the bots to learn even more effective movement strategies.

While the technology is still new, it has been tested using computer simulations, and the team has developed a small palm-sized robot that runs the algorithm on a tiny chip. The next step is testing the AI on further types of flow disturbances it may encounter in the ocean to assess effectiveness. The team is placing a custom-built drone dubbed the CARL-Bot in a two-story-tall water tank on Caltech’s campus and taught to swim. Ultimately the research could be useful in a wide range of applications: ocean surveying, monitoring of deep-sea animal communities, and, out of the ocean and into the air currents, tasks such as drone-based inspection and delivery and weather balloon station keeping.

Source link

%d bloggers like this: