Image to Trajectory Data Collection for Annotation-Free 6D Cutting Pose Learning in Grape Harvesting Robots — Laboratory Teleoperation Dataset

This dataset contains raw video recordings and robot joint trajectories of teleoperated grape-cutting demonstrations, collected in a controlled indoor laboratory environment (artificial grape clusters and foliage, artificial lighting). A Universal Robots UR10e arm, fitted with an effector carrying an onboard GoPro camera (155° fisheye lens), was piloted remotely by an operator watching the live camera feed toward a target grape cluster — the same teleoperation protocol used for our vineyard (field) data collection. Each of the 672 sequences (541 train / 87 validation / 44 test) provides a video.mp4 recording and a trajectory.csv log of UR10e joint angles. A provided Python script (process_teleop_session.py) converts each sequence into training-ready image frames and 6D cutting-pose ground truth (remaining translation and rotation to the demonstrated cut point), computed via forward kinematics — no manual per-frame annotation is required. A minimal PyTorch dataset loader (load_dataset.py) is included as a usage example. Intended for training and evaluating vision-based 6D pose estimation / end-effector control models for autonomous grape-harvesting robots.