Part 7/10:
One of the foundations of Helix's success lies in its innovative data collection method. The model was trained on a high-quality dataset derived from diverse teleoperated behaviors, amounting to 500 hours of recorded actions. This data collection approach not only ensures a variety of training scenarios but also integrates natural language descriptions that correspond with the visual and motor tasks performed by the robots.
This method reflects an evolution in training paradigms, seeking to overcome the limits of traditional data collection that often constrained the adaptability of AI models to new environments.