The database contains data obtained from executions of a peg-in-hole task (using the Cranfiled benchmark). We used a Franka Emika Research 3 robot, an ATI Nano25 force/torque sensor, and an Intel Realsense D435 RGB-D camera. The database is structured by its offset from the in-hole position. The folders are organized according to the following structure: In the first folder level, the positional offset is specified. The orientational (rpy) offset is defined in the second folder level. At the third level, we have two folders: - one for the the trajectory, that led to the occurrence of the error, and - one for the human-demonstrated trajectories, that show possible correction strategies. Each entry is formatted as MATLAB data file (`.mat`) with a states structure, that contains: - reference and measured target pose in Cartesian space (positions and quaternions; sampled at 100Hz), - raw force-torque sensor data and force-torque data measured in the robot-wrist (sampled at 100Hz), - RGB image sequence (sampled at 10Hz), - point-cloud sequence (sampled at 10Hz). We use the following labeling for file names. The first part denotes positional offset: - x10 ... +10mm in $x$-direction from target position - xm10 ... -10mm in $x$-direction from target position - y10 ... +10mm in $y$-direction from target position - ym10 ... -10mm in $y$-direction from target position The second part denotes orientation offset: - x10 ... +10 degrees around $x$-axis - y10 ... +10 degrees around $y$-axis - z10 ... +10 degrees around $z$-axis The ending `.cor.N` stands for the $N$-th correction trajectories, and `.error` for the trajectory where an error occurred. Example: The file `x0y10_x0y0z0.cor.1.mat` contains data from a correction trajectory where at the onset of the error, the peg is displaced from the target pose for 0 mm in $x$-direction, 10 mm in $y$-direction and there was no orientational offset.