Researchers propose SSC for robot task labeling

Researchers from the University of Hong Kong have developed the Structured Subtask Chain (SSC), a novel approach to labeling bimanual robotic manipulation tasks that bridges the gap between unstructured natural language and rigid template formats. Published as arXiv:2608.04425, the system represents demonstrations as sequences of Structured Subtask Template (SST) entries, each containing core action components (subject, predicate, object), conditional modifiers, base-motion fields, and after-state scene graphs.

The innovation lies in SSC's ability to render these structured representations as natural language while maintaining verifiability through four state-transition rules. When deployed on BEHAVIOR-1K's 50 tasks (3 episodes per task, 2,357 annotated actions), the system identified labeling anomalies across 13 state-of-the-art vision-language models tested as verifiers. SSC's query resolution cascade also enables automatic completion of underspecified fields, significantly improving annotation consistency for long-horizon robotic tasks.