Robots Are Learning Welding and Tile Laying on Construction Sites

Vision becomes part of the control loop

I see this not as another flashy video of a robotic arm, but as a concrete technical shift: the robot is no longer merely replaying a pre-recorded movement. A Vision IA publication demonstrates robotic welding, while an OVACEN overview examines tile laying. In both cases, the core idea is the same: a camera determines the object’s actual position, and the system then corrects its motion.

For welding, computer vision detects edges and the seam line, converts pixel coordinates into a trajectory, and sends corrections to the controller. The cited technical studies report an error of about 0.008 mm for ANFIS-based control. Another paper claims localization accuracy of up to 99.998% across nearly 96,000 operations, with precision of 1.00 and recall above 0.99.

Tile laying involves similar mathematics but different physics. The system must estimate the tile’s position and orientation, compensate for a limited field of view, and then inspect the joint and surface level after placement. Edge detection, Hough transforms, finite-state machines and optical measurements are used for this purpose.

Commercial materials claim grout-joint errors of up to ±0.1 mm, surface deviation within 1 mm, and work speeds five to six times faster. As of this review in September 2026, I would treat these figures as supplier claims rather than universal results for every jobsite. Uneven substrates, dust, glare, different material batches and unstable geometry quickly bring a robot from a demo back into reality.

What is being automated is a controllable slice of work, not a profession

The main change is that specialized manual operations are becoming suitable for closed-loop automation. A camera detects deviation, a sensor confirms contact or force, a planner recalculates the motion, and quality control happens immediately rather than after the section is complete.

This is real progress, but it is not proof of full autonomy. The first things I would examine are the share of manual preparation, the frequency of stops, recalibration requirements and resilience to rare defects. Laboratory accuracy says little about throughput if a person must constantly rescue the cycle.

Welders and tile setters are not disappearing yet, but the boundary between craft work and operator work is shifting. The most interesting question is no longer whether a robot can perform a perfect pass, but how much imperfect construction it can handle without intervention.

We previously examined how embodied AI demonstrations can blur the line between an impressive prototype and a deployable system. That distinction also matters as robots move into skilled trades, where physical reliability and operational design determine whether automation works.