Guided by advanced models, the systems Jaime Fernández Fisac designs will navigate simulated street traffic, maneuver gusts in synchronized flight, and lead human dance partners in a back spot turn. But the models are not the real world. In the gap between theory and broad use that widely benefits society, lurk the dangers of overconfidence. And Fisac insists these robots be safe.

“No matter how carefully you design your model, and no matter how much data you feed your machine learning algorithm to train a very accurate model, at the end of the day this gap will exist,” said Fisac, an assistant professor of electrical and computer engineering who joined the faculty in fall 2020. “And it’s going to show up at the least convenient time.”

If he can train the systems to reason about those gaps, Fisac believes robotics will be in a far safer position. He combines game theory (a field rich in Princeton history) and cognitive science with reinforcement learning, an AI technique that trains the models to reason through millions of variations of a task.

Dance provides a neat case study as robots’ sophisticated algorithms learn how to read human behavior in ever finer detail. And they will eventually be able to predict all likely outcomes for a given scenario, moment by moment. That could be arm in arm, dancing the bolero, or zipping along the interstate in heavy traffic.

But Fisac zeroes in on the qualifier — likely. It helps him flag the chasm between the plausible and the real. That sparse, intractable outer limit of the probabilistic long tail. That’s where we find real surprise. And Fisac says to guarantee safety, robots need to respond dynamically to their own failing predictions, anticipating and avoiding no-win situations.

“There’s this mutual dependence between your decisions and the decisions of others that makes or breaks your safety,” he said. “It really is a dance. And it’s probably the most important form of uncertainty that we’re going to need to grapple with in modern robotics.”