Encouraged by recent progress in AI, some startups are focused on having robots learn in simulation how to perform fiendishly difficult tasks like grasping irregular objects, technology that could eventually help automate much ecommerce and logistics work. This often uses an AI approach called reinforcement learning, which involves an algorithm experimenting and learning, from positive feedback, how to achieve a specific goal.
“This is definitely the way to go,” says Ding Zhao, a professor at Carnegie Mellon University who focuses on AI and digital simulations. Zhao says simulations are crucial to using AI for industrial applications, partly because it is impossible to run machines through millions of cycles to gather training data. In addition, he says, it’s important for machine-learning models to learn by experimenting with unsafe situations, such as two robots colliding, which cannot be done with real hardware. “Machine learning is data-hungry, and collecting it in the real world is expensive and risky,” he says.