Researchers just taught robots to predict your every move


In a few years time the Droids from Star Wars are going to seem like relics. Today’s robots might be better suited for sewing clothes and building cars, but tomorrow’s could be as indispensable and ubiquitous as our smartphones are.

A group of researchers in Europe recently published a white paper unveiling their experiments in teaching robots to anticipate human movements. The team’s work, to create “robots that can predict human actions and intent, and understand human non-verbal cues,” could pave the way for innumerable advances in the field.

The researchers focused on combining previous research teaching AI to understand human gaze and pose, both of which are crucial for robots to understand in order for them to work with humans.

According to the team:

Situations where fast cooperation is essential, for example cooperative assembly, require the understanding of subtle non-verbal cues about the human intention and future action. In these scenarios it is not enough to merely recognize the current action. Instead, it is fundamental to predict actions and anticipate the intent in order to guarantee seamless cooperation.

Using black box AI in the form of a recurrent neural network (RNN), the machines learn to determine the “intent” of a person based on where they’re looking (gaze) and the pose of their body. Whenever it gets things right, it transfers its newfound “knowledge” through the various layers in its RNN where the data is then normalized.