SPARC: Reliable Spatial Annotations from Robot Demonstrations at Scale
Researchers have developed SPARC, a new framework for generating reliable spatial annotations from robot demonstrations. SPARC assigns a confidence score to each annotation, addressing the issue of noisy labels in existing automated pipelines. This framework leverages the inherent spatio-temporal structure of robot tasks to improve annotation quality and retain more useful data. SPARC has demonstrated state-of-the-art results on object-grounding and pointing benchmarks, and policies trained with its annotations perform better in real-world cluttered environments. AI
IMPACT Enhances the reliability of data used for training embodied foundation models and robot policies.