Autonomous Human-Robot Interaction with a Model Trained on Operator Data

Here is a robot that can imitate its operator for easy control. Most of the time, teleoperated robots are designed to perform expressive interactions with humans. In this case, it is done autonomously with a model trained on operator data. As the researchers explain:

Our approach learns to predict continuous operator commands through a diffusion process and discrete commands through a classifier, all unified within a single transformer architecture.

This video shows how it works.

[HT]