Depaletization Sequence Powered by Computer Vision


Ciseo specializes in the development of production equipment and specific machines for production lines in the pharmaceutical industry. The company encountered a problem to automate depalletization of boxes of syringes stored in crates. These boxes have vertical sides ending with a horizontal area facing outwards, which overlap between boxes in the crates. As a result, automatic depalletizing by a robot, using a sequential order, is almost impossible without spilling the contents of the boxes. 

This is why Ciseo contacted Sagacify to help them develop an AI system that lets a robotic arm determine which is the next box to be taken out of the crate in order to avoid breakage.


Sagacify developed an AI algorithm to analyze the images taken by a camera placed on the robotic arm. The AI model allows to determine which box should be removed next, i.e the one with all sides visible without overlap, identify its coordinates (x,y,z), and send the information to the robotic arm in order to empty the crate without breaking.

To cope with the challenges explained above, the problem was broken down in different stages: 

  • Detection of the areas of interest for each of the boxes in the crate on a reduced resolution image. The symmetry of the problem with the boxes arranged in 4x4 grids made it easier for the model to detect these areas. 
  • Detection of the corners of the different boxes on a zoom of the areas of interest detected in stage 1, this time using the full resolution of the image. 
  • Identify boxes whose corners are all free, i.e. no overlap with another corner. To this end, boxes are scored to assess which box has the highest probability of being the one to be extracted. 
  • The coordinates of the box with the highest probability to be free are then sent to the robotic arm so that it can be removed from the crate. 
  • The operation is then repeated until the crate is completely empty.


  • Custom model to create a safe depaletization sequence from 2D images on a robotic arm
  • Deployment of a custom labellisation environment 
  • Model deployed on the edge, on a local compute unit located on the robotic arm

The developed model is able to correctly detect a box that can be picked up by the robot, with an accuracy of 92.5%. In most cases, several boxes have all their corners free at the same time which means that the probability to remove a box without breaking any syringes is very high.

Related cases