More capable machines. More productive work.
We bring perception and AI into industrial robotics. From inspecting a changing part to assisting a production task, we connect intelligence to the cell where it creates value.
The perspective
Automation that can handle variation.
Fixed routines work well until the material, product or context changes. AI can help machines interpret that variation. Welf engineers the connection between perception, production knowledge and the task on the line.
Inspection workflow / example
From camera frame to a reviewable result.
A proposed workflow for a single inspection station. The first objective is to establish reliability across real production variation.
Capture the context
Bind the image to the part ID, revision, camera settings and station. A good prediction attached to the wrong part is still a failed inspection.
Prepare the finding
Locate the suspected defect and attach the relevant inspection criterion. Flag unfamiliar parts and uncertain images for review rather than forcing a pass or fail.
Return the outcome
Quality staff resolve flagged cases. Record the approved disposition in the quality system and retain the evidence needed to investigate later.
A closer look
What goes into the work.
Explore the questions that matter for your project.
Where does variation
slow the line?
Visual inspection
Check parts against defined defect classes. Keep the image, part variant and inspection criteria with the result so quality staff can investigate uncertainty and false rejects.
Operator assistance
Retrieve the setup instructions, tooling records and approved checks for the machine and part revision in front of the operator. An uncertain match should trigger a question.
Robotic task support
Assess perception or planning for a bounded cell operation. Test the range of parts, poses and environmental conditions before allowing any command into the automation layer.
Fit into the cell’s
existing safety design.
Perception and language models do not replace safety functions. Your automation specialists define permitted commands, interlocks and safe states. Machinery changes require the appropriate engineering assessment.
Test the conditions
Lighting changes, reflections, occlusion, new finishes and camera movement belong in the test set. Check performance by defect class and part variant.
Plan the interruption
Define what happens when a camera fails, the network drops or an operator intervenes. A restart must not repeat a physical action that already occurred.
Choose the level of authority.
| Stage | The system can | Evidence needed |
|---|---|---|
| Observe | Run alongside the existing inspection process. | Matched cases, labeled defects and a comparison with the current method. |
| Assist | Show findings to an operator for disposition. | Useful explanations, acceptable review effort and reliable uncertainty routing. |
| Act within limits | Issue explicitly permitted actions through the agreed controller interface. | Cell-specific acceptance, tested recovery and approval by the responsible engineering team. |
Before a robotics project.
Do we need new hardware?
Not necessarily. We first inspect camera coverage, controller interfaces, cycle time and the existing inspection process. New sensing or compute is considered only where the task needs it.
What should the first evaluation include?
Representative good and rejected parts, uncommon defects and normal shifts in production conditions. Measure missed defects and false rejects separately, together with cycle time and the number of cases sent to an operator.
What do we bring to the first discussion?
The inspection plan, sample images or parts, the cell layout, available interfaces and a named quality or automation owner. Share only material you are authorized to disclose.
Bring intelligence onto the line.
Bring the part, the inspection criterion and the variation your team struggles with. Start with one task whose result can be tested.
Start a conversation