BlockManip: A MuJoCo-Simulated Robotic Manipulation Dataset

BlockManip is a MuJoCo-simulated robotic manipulation dataset designed for vision-based pick-and-place and block-stacking evaluation. Each episode records one manipulation trial from a fixed external camera view and contains four phase-level RGB observations: init, grasped, preplace, and final. This release contains 2,750 episodes, including 2,500 training episodes and 250 evaluation episodes, for a total of 11,000 RGB frames. Each episode includes four RGB images, one episode-level label file, and one trajectory trace file. The release includes RGB observations, per-episode labels, trajectory traces, train/eval splits, metadata, and SHA-256 checksums for integrity verification. Depth maps and segmentation masks are not included in this release.