Automated picking2018-08-13T14:02:24+00:00

Automated picking of varying parcels

The Challenge

For years (random) bin picking robots are programmed to successfully pick known and often identical items. However, in e-fulfillment one needs to handle hundreds of thousands or even millions of SKUs which also frequently change. This cannot be solved by conventional programming techniques of teaching in each SKU or 3D CAD model. Since winning the Amazon Picking Challenge in 2016, Fizyr successfully applies Artificial Intelligence on machine vision to enable robots to cope with variation. Our latest deep learning algorithms allow picking of unknown parcels from a cluttered pile!

Fizyr is #1 in picking of unknown objects and tightly packed boxes

Singulation of parcels for handlers like UPS, DHL and FedEx is hard. All jostling parcels are unknown and positioned in bulk. The parcels can weigh up to 30 kg and the number of angles, positions, and orientations is infinite.

Based on the segmentation of all objects the AI software differentiates boxes between bags, cylinders, flasks, etc. defining the best grasp strategy. It will also handle singulation of tightly packed boxes.

Using machine learning and gripper-vision coordination we almost instantaneously propose 6 DoF grasp poses enabling the robot to pick the items one by one without waiting on guidance.

Pick the unknown from anywhere

Fizyr AI software allows picking from a box, tote, tray, shelve, accumulation chute, cabinet, container, roll cage or fast-moving conveyor belt. The objects may vary in material, shape, size and color like boxes, clothing, bottles, jiffy bags, plastic bags, padded mailers, flasks, cylinders, apparel, etc.

Winner of the prestigious Amazon Picking Challenge

As robotic integrator we participated once in 2016 and won both the picking and stowing challenge. This led to our pivot to become a computer / machine vision company and fully focus on developing AI based item and parcel picking solutions for global logistic integrators.

Training of the neural network

The last few years we trained our deep neural network effectively with millions of images. This means our generic algorithm already has high quality by default with new camera’s and unknown items and parcels within completely new environments. However, we can further improve by training once on a set of images of the actual situation of the client. Once installed we can continue the training and further optimize the picking performance and robustness. Since Fizyr performs this globally, all users can benefit from exchanging local learnings.

Goods-to-person and Automated Storage and Retrieval Systems (ASRS)

The ideal business cases are goods-to-men situations with a static picking station. Think of ASRS with conveyer belts, shuttles, AGV’s or AMR’s. In 2017 Fizyr performed the picking of varying and unknown items from an AutoStore. Our machine vision software can also be used on mobile platforms to perform classifications and manipulations.

Most Fizyr clients are global robotics integrators in logistics

We develop and tailor our software for our partners and create a comfortable long-term Win-Win situation.

Three type of Fizyr partners:

  1. Suppliers / integrators of Material Handling Equipment (MHE)
  2. Robot vendors and integrators
  3. The big end-users in online retail, warehousing, parcel distribution and courier / express

Inspection and Quality Control

We can train the same algorithm to find unknown defects and perform grading of quality. An example use case is preventing broken boxes or bags to be placed on the sorter.


Fizyr has trained AI software to enable automated picking of unknown items and parcels. Leading global players in logistics apply it on robots for item picking and parcel handling.


Motion / path planning & collision avoidance

Having been a robotics integrator, Fizyr still performs Proof of Concepts, we are able to help you on robot guidance and grasp planning. However most of our clients are global leaders in their field and just want all potential 6DOF grasp poses with our proposal of preference.

Maximum size, weight, distance

These maximum size, weight and distance depends on the hardware you choose.

Number of picks per hour

Our software can provide the 6DOF grasp locations of all items within 200 milliseconds, many times faster than a robot can perform pick-and-place.

Item classification

Objects can be classified on shape, color or size to be separated and handled differently.

Unloading, depalletizing and devanning of multi SKU’s or parcels

The software is trained to cope with variation, disordered loading and other unknown situations, so it can cope with random pallets and containers.

Feeding sorters, packaging machines and conveyor induction

After picking, parcel orientation can be adjusted and placed with high accuracy.

Hardware independent

We believe you should have the flexibility to choose the best tools for your application. Fizyr ensures the use of the best picking software. Fizyr picking software operates with off the shelf hardware, from cameras to robots, grippers and PC’s.

100% hardware independent

Sensors, 3D cameras and lasers

Fizyr uses the best 2D and 3D (including RGBD) cameras and lasers available to cope with changing and poor light conditions, reflective surfaces, overlapping products and other local challenges.

Choose your robot

Fizyr picking software works with all common robots including ABB, ABI, Fanuc, Kuka, Mitsubishi Electric, Omron, Stäubli, Universal Robots and Yaskawa Motoman. Based on the requirements you can also choose cobots or very fast delta robots.

Any gripper / end-effector

Fizyr software can work with many types of grippers including different configurations like f.i. grabbing from different angles or variations in the diameter of vacuum griper suction cups. Based on the (custom) gripper at hand, the grasp strategy can be optimized.
NOTE: Although Fizyr is fully focused on developing software we made a new versatile vacuum gripper because they we’re not existing yet. You can simply 3D print them and license the designs.

Standard Linux PC with a GPU

You can use a normal PC with a standard GPU, like NVIDIA, running under Linux. Since the algorithm only requires a few hundred milliseconds one pc might provide grasp locations for multiple robots. By using a TCP/IP connected camera you have the freedom to locate the PC on your ideal location.