Which factory tasks robots still can't do: deburring, harnesses, sewing

What makes a factory task hard for a robot#

A factory task is hard for a robot when the material deforms, the parts vary from one to the next, or the work depends on controlled contact and feel. Deburring, wire harness assembly and sewing combine all three, which is why they remain largely manual in many plants even where welding, painting and palletizing are automated.

Robots do best with rigid parts, fixed positions and long runs. Each departure from that adds cost: vision to find the part, force control to touch it, custom tooling to hold it, and programming time for every variant. In high-mix, low-volume plants, that programming time often outweighs the labor saved.

Physical AI aims to close the gap by learning from examples rather than explicit programs. That shifts the question from how to program a task to what examples of skilled people doing it actually exist.

Hard tasks, why they resist automation, and the records that help#

Hard-to-automate manufacturing tasks share root causes, and the records that help AI developers are the ones that capture how skilled people handle the variation. Most plants already keep some of these records for quality reasons.

Deburring: variation you can feel#

Deburring resists automation because the burr is a byproduct, not a designed feature. Its size and location change with tool wear, material lot, feed rate and the part's orientation in the fixture, so a robot following a fixed path either misses burrs or cuts into the edge.

Skilled operators adjust pressure and angle by feel and check the edge with a fingertip or a loupe. Robotic deburring cells work well on consistent parts with compliant tools and force control, and struggle on high-mix work and castings. Records that capture what operators actually do, such as rework notes and edge-related NCRs tied to tool wear, are what learning systems lack.

Wire harnesses: the deformable-object problem#

Wire harness assembly is hard because wires are deformable objects that bend, twist and keep memory from the spool. Routing a bundle through clips on a form board, inserting terminals into connector cavities and wrapping branches with tape all mean handling something that changes shape as it moves.

Harness work is also high-mix: many plants build a large number of harness variants in small lots. Records that help include form board layouts, build instructions, continuity and hipot test results, and defect logs for miswires, back-outs and damaged insulation, linked to the build that produced them.

Sewing: fabric does not stay where you put it#

Sewing is hard because fabric stretches, slips and puckers, and an operator constantly realigns plies as the machine feeds them. Automated sewing handles some long, repeatable seams, but curved seams, mixed materials and three-dimensional assemblies such as seat covers still rely on people.

Patterns, seam specifications, operation breakdowns, defect codes by operation and training material are records sewing plants already keep. Linked to outcomes such as rework and customer returns, they show which steps carry the most judgment.

Material handling around the sewing head is often the harder part. Picking a single ply from a stack, flipping a panel or easing fabric around a curve are the steps where automated cells stall, and they are rarely written down in any instruction because operators simply do them.

What physical AI developers look for in factory records#

Physical AI developers look for examples of real work linked to outcomes: what the person did, under what conditions, and whether the result passed. Video of skilled work is one form. Build AI's Egocentric-10K dataset card, for example, describes head-mounted video collected exclusively in real factories and released under an open license.

Most manufacturers hold more modest but still useful records. These gain value when they connect a task to its result.

Video of employees raises privacy and labor questions. Notice and consent, collective bargaining agreements and biometric privacy laws such as Illinois BIPA may apply, and customer parts visible on camera can be confidential. Review these with counsel before recording anything new.

• 1Work instructions with revision history and the defects that prompted each change.
• 2Rework and repair records with cause codes and the operator's notes.
• 3Inspection and test results tied to work orders and stations.
• 4Training videos and photos, reviewed for people, customer parts and logos.
• 5Line stoppage and delay notes that show where tasks slow down.

Illustrative example: a seating manufacturer reviews its records#

Illustrative: a fictional manufacturer of commercial vehicle seating runs sewing, foam and assembly lines, including small wire harnesses for powered seats. Its records include digital work instructions, defect codes by station, rework logs and training videos made for new hires.

The CEO asks whether any of this matters to robotics developers. The team inventories existing records first and films nothing new. Defect codes and rework logs link cleanly to work orders. Training videos show employees' faces and a customer's branded seat covers, so they are set aside pending consent and customer review.

The company submits a description of the record inventory, not the files, for a fit check, with the defect and work instruction history as the leading candidate.

How SourceX approaches physical-world records#

SourceX evaluates physical-world records with the SourceX Enterprise Data Value Framework, where human-generated signal, domain expertise and uniqueness raise value, and privacy burden and preparation cost reduce net value; that is why video of people needs careful review before it is considered. Work moves through the SourceX five-step transaction, starting with a metadata-only fit check, and the manufacturer approves each step.