Robot manipulation dataset with structured chain-of-thought annotations, derived from five heterogeneous data sources covering real-world kitchen manipulation, tabletop grasping, and simulation environments.
Each sample pairs a robot-perspective scene image with a natural language task instruction and three-tier structured annotations including object attributes, spatial relations, and atomic action subgoal decomposition.
115,158… See the full description on the dataset page:
https://huggingface.co/datasets/mingyueyue/Bridge-CoT2.
