Therealworld ai is an immersive virtual training ground that may help AI learn to behave more humanly. It could also help developers test and refine their apps before releasing them to consumers.
For instance, one of the reasons why self-driving cars have failed to take off in the market is that they are designed to drive in lab conditions — which aren’t always identical to the real world. Creating realistic simulations like this will make it easier for engineers to figure out how to integrate their AI into the real world, Mottaghi says.
Innovation Unleashed: How The Real World .ai is Shaping Our Future
Many previous training environments for AI have taken their cue from commercial video games, including the first-person shooter “Doom” and open-world driving game “Grand Theft Auto.” But these are limited in the types of actions that can be performed, which makes it hard to train an agent to do more than fire a weapon or move from one point in the world to another.
This new VR environment, which Mottaghi has been working on since 2016, is designed to mimic more of the everyday tasks that people do, such as walking through a house or using a microwave. It can be populated with objects that can be interacted with, such as sliceable apples and empty or made (or unmade) beds, and it has the added bonus of being able to simulate physics. In the future, the team hopes to add features such as non-rigid body physics that would allow agents to practice folding a sheet or manipulating a pillow.