Fizyr supplies vision AI for robotic picking and mixed-material handling

Fizyr develops vision AI that helps industrial robots interpret mixed scenes and choose useful handling actions. Its product is delivered through automation projects and integration partners, making it relevant when a robot cell needs perception for variable parcels, pallets or other items.

01 / ProductA vision component inside a larger automation system

The Fizyr OS product page describes perception for detection, picking, placement, singulation and pallet handling. The software provides information that a robotic system can act on, while the surrounding cell supplies the cameras, robot, gripper, conveyors and controls. This is a useful distinction for buyers: improving visual decisions can unlock an application without making the software a complete warehouse system.

The company history traces Fizyr to Delft Robotics and a 2017 rebrand as the business focused on vision AI. The current company presents an integrator-oriented offer rather than a catalogue of its own robot arms. That explains why the implementation partner is central to the buying conversation, especially for customers who need one party to deliver the complete physical process.

On its technology page, Fizyr describes identifying items in a scene, prioritizing handling choices and directing the robot control system with grasp-related information. It also describes inference on industrial servers and cloud use for monitoring and improved learning. These are vendor descriptions of the execution model, not a verified account of private implementation details or a guarantee about every installed configuration.

02 / AudienceUse it where variable scenes make fixed picking rules brittle

The pick-and-place material discusses 2D and 3D capture, segmentation, object characteristics and partially obscured items. Those capabilities are relevant to a parcel stream where the robot must decide which item can be picked next and how to approach it. A factory presenting every part in a precise fixture may have a different perception problem and should compare simpler alternatives.

The pallet-handling page focuses on mixed package types and stacks disturbed during transport. Here the challenge includes the arrangement of neighboring items, not just recognizing a single object. A suitable project has a defined output, such as transferring individual parcels onto a conveyor, and enough representative material to test whether useful grasp choices remain available through the whole pallet.

Integrators are a particularly clear audience because they can connect perception to motion and process controls. An end customer should establish which integrator owns the complete result, and whether Fizyr’s role includes configuration, optimization and continued support for that project. The partner page describes collaboration with certified integration partners and component providers, but the responsibility split still needs to be explicit in an individual proposal.

03 / WorkflowA proposed mixed-parcel cell separates seeing, grasping and completing the transfer

Consider a proposed cell that removes parcels from an irregular pile and places them onto a conveyor. The first step is to define the allowed parcel types and the output condition. An item arriving on the conveyor in the wrong orientation, too close to another item or with a damaged package may not count as a useful completed transfer. This is an evaluation example, not a trial performed for this article.

The team then selects camera positions and representative test scenes. The aim is to expose the conditions that matter in the facility: overlap, reflective wrapping, soft surfaces, clutter and changing pile height. A strong test set includes cases where the correct outcome is to decline a pick or request intervention, not only cases in which a grasp should succeed.

Fizyr’s technology description explains a sequence from scene segmentation to prioritized picking and guidance for the robot controller. In the proposed cell, that output must be connected to a feasible physical motion and a suitable gripper. A vision system can identify a promising contact area while the actual tool still struggles to seal, hold or release the item. The evaluation should keep those causes distinct.

After each transfer, verify the outcome and observe the next scene. Removing one parcel can uncover another or destabilize its neighbors. Fizyr’s product material includes placement verification as part of pick-and-place capability. The implementation should make clear how that information affects the next action, how failed transfers are recorded and when a person is asked to intervene.

The final test should include the receiving conveyor’s behavior. If the destination stops, the robot needs a defined response that preserves the process state. Tracking complete, accepted transfers through the cell gives a more useful performance measure than counting visual detections or issuing grasp coordinates. It also helps identify whether the next improvement belongs in perception, tooling, motion or downstream equipment.

04 / PricingThe commercial offer is scoped around an application and partner relationship

The homepage describes a pilot path from a prototype through performance measurement to production preparation. The contact page directs prospective customers to a project conversation. We found no public numerical license tariff on the pages reviewed. The invitation to explore a pilot should not be interpreted as a published free-production entitlement or a standard contract with known limits.

The product page combines software with consulting, implementation participation, training and ongoing support. Those services are part of the proposition, but the public description does not establish the exact hardware package, service hours or update terms for every customer. A quotation should identify which work belongs to Fizyr and which belongs to the cell integrator.

For an integrator building a repeatable product, the commercial question also includes reuse. Confirm whether the agreement covers one cell, a named customer, multiple deployments or a broader product family. We do not infer those rights from the software’s hardware-agnostic positioning. Technical compatibility and commercial permission to distribute or operate software are separate questions.

05 / DistinctionsSpecialized perception can preserve choice in the rest of the cell

Fizyr’s partner model emphasizes combining its vision with different robots, cameras, grippers and cell components. That can be useful when an integrator already knows the mechanical design and wants a specialized perception layer. The practical benefit is choice within a supported integration, not proof that every combination of industrial hardware will work without adaptation.

Cognex is an adjacent comparison for teams deciding whether their core need is inspection, measurement or robot-guidance perception. Ambi Robotics offers a comparison around more packaged logistics manipulation systems. Fizyr’s appeal is strongest when the buyer or integrator wants to retain responsibility for the overall cell while sourcing a specific vision capability.

A September 2026 KLM announcement describes work with KLM, DERO Group and Viroteq on mixed air-cargo pallet handling. It explicitly calls the work a proof of concept, not a production deployment. That makes it useful evidence of technical exploration, while preventing a reader from assuming the same workflow is already operating as a fully deployed service at KLM.

The broader lesson from that example is that feasibility and production readiness answer different questions. A demonstration can establish that a difficult mixed scene can be handled at all. A production decision additionally needs sustained operation, recovery procedures, maintenance ownership and an agreed treatment of the items that remain outside scope. The project should name which question each phase is intended to resolve.

06 / QuestionsResolve the meaning of accuracy and the ownership of updates

Fizyr’s public pages contain strong accuracy and latency claims, with different figures and scopes. We have not adopted those claims as benchmarks or ranked the company using them. A buyer should request the definition behind the relevant measure: detection accuracy, grasp success, successful placement and completed cell throughput are not interchangeable. Ask for evidence using the package mix and timing boundaries that matter to the proposed application.

The edge-and-cloud description also leaves project-specific questions about data handling and offline behavior. Establish which images or operational data leave the site, how long they are retained and what functions remain available if connectivity is interrupted. The public architecture describes a relationship between local execution and learning; it does not supply a complete deployment policy for every customer.

Finally, clarify how improvements reach an installed cell. A model update may help with a new item class while still requiring validation against the existing workload. The integrator and software provider should agree who tests changes, who authorizes rollout and how to return to a previous working version. This is especially relevant when the same cell design is replicated across customers with different material mixes.

07 / DecisionChoose Fizyr when perception is the component that needs specialist work

Fizyr is worth considering for a well-defined handling application where variable scenes make perception a limiting factor and an integrator can own the physical system around it. Start with representative materials, measure completed transfers and establish the commercial and support boundaries. Treat proof-of-concept examples as evidence about feasibility, while asking for separate evidence about the production arrangement you intend to buy.