Groundlight has released an open-source ROS package designed to simplify the integration of computer vision into robotic systems. Announced on October 14, 2024, the package allows developers using ROS2 to incorporate visual intelligence more efficiently into their projects. It combines machine learning with human oversight to help robots better manage unstructured environments.
Traditional computer vision workflows often require extensive data collection, labeling, and model refinement, which can be time-consuming. Groundlight’s package aims to streamline this process by enabling customized edge models that operate locally on robots, with cloud-based training and human intervention. When the system encounters unfamiliar situations, it pauses and waits for human input, which is typically received quickly. This feedback is then incorporated into the model to refine its performance.
Leo Dirac, CTO of Groundlight, stated that the package provides a faster, more reliable alternative to existing methods of visual processing for embodied AI. He explained that combining fast edge models with human oversight enhances the system’s efficiency in robotic applications.
The ROS package allows developers to ask questions about images in natural language. Responses are handled automatically by the machine learning model when confidence is high, with low-confidence cases being referred to human reviewers. This continuous feedback loop aims to improve model performance over time without requiring manual retraining.
