Sensing & Adaptive Fabrication

Most fabrication is open-loop: the machine runs a toolpath that assumes the stock is exactly where and what the model said. For a fresh sheet of acrylic on a laser bed, that assumption holds and the part comes out perfect. For real material with a mind of its own, it fails. Timber warps and each board differs; reclaimed and rough-sawn stock has no true datum; stone and rammed earth are irregular by nature; clay slumps under its own weight as you print it; a weld pool wanders; a placed brick is never exactly where the plan drew it. A blind path meets that reality and gouges where the surface is high, skips where it is low, and stacks up error as an assembly grows.

The fix is to stop assuming and start measuring - to close the loop between the machine and the world in front of it. A machine that senses the actual material and adapts its path to what it finds is adaptive fabrication. This is what lets robots work with cheap, variable, sustainable material instead of only pristine stock; it is what makes a big multi-step assembly stay true; and it is, more than raw precision, what makes robots genuinely robust on a real site rather than only in a clean lab.

A machine can perceive its work in several ways, and each suits a different problem. Laser line scanners and structured-light sensors sweep a stripe or pattern across a surface and reconstruct its 3D profile - ideal for reading an irregular timber face or a rough stone block before machining. Vision - a camera plus computer-vision software - finds edges, fiducial markers and features to locate a part, track a seam, or verify placement; a camera watching an FDM print can spot a failed first layer or spaghetti. Touch probing taps the surface to find a precise datum - the CNC world's classic 'touch-off' that sets Z-zero on the real top of a warped board rather than a nominal one. Force and torque sensing at the wrist lets a robot feel contact: press until a defined force, follow a surface at constant pressure, or stop when it meets resistance - essential for polishing, sanding, assembly and any task where position alone is not enough.

Real cells often combine them: a scan to plan, vision to track, force to finish. The common thread is converting the physical state - profile, position, contact - into numbers the controller can act on, at a rate fast enough to matter.

A concrete pairing shows how they layer. To robotically cut a mortise into a reclaimed oak beam, a robot might first sweep a line scanner across the beam to build a mesh of its real, cupped face; use vision to read a printed marker and lock the beam's exact position in the cell; then drive the milling spindle with a force limit so that if it meets a hidden knot or nail it eases off rather than snapping the cutter. One task, three senses, each answering a different unknown - what shape, where, and how hard. That layering is typical of serious adaptive work.

Adaptation comes in two tempos. Scan-then-plan measures first, then computes a bespoke path for this piece of stock: laser-scan a warped oak beam, build a mesh, and generate a joint or a surfacing pass offset a constant depth from the real surface, so the cut depth stays uniform even though the board is not flat. Robotic timber work at labs like ETH Zurich and ICD Stuttgart leans on exactly this - each irregular or reclaimed member scanned and its path derived, so non-standard material becomes usable. It is robust and easy to verify because the plan is fixed before the tool moves.

Real-time adaptation adjusts during the operation from live feedback. Seam tracking nudges a welding or extrusion head to follow a joint that drifts. Adaptive milling eases the feed rate when spindle load climbs, protecting the tool. Flow and speed on a clay or concrete printer trim to how the last layer actually settled. In assembly, visual servoing guides a gripper onto a target it keeps in view, correcting as it closes. Real-time is more powerful and less forgiving: latency, noise and stability all bite, and a badly tuned loop can chase its own errors. Most practical work scans-then-plans, and reserves true real-time control for tasks where the material genuinely will not sit still.

The choice is really about where the uncertainty lives. If the material is irregular but stable once placed - a warped board that stays warped - scan-then-plan captures that once and is done. If the uncertainty is live - a seam that wanders as you weld, clay that keeps settling, a part still being positioned - only real-time feedback can track it. Many mature cells hybridise: scan to plan the bulk of the job, then let a fast local loop handle the last, moving unknowns. Knowing which kind of uncertainty you face is half of designing the process.

Underneath every example is one loop: sense, adjust, fabricate, repeating. Sense the real state; compare it to what the plan expected; adjust the command to close the gap; act; and sense again. This is elementary control theory, and two ideas from it are worth carrying. Feedback corrects after measuring the error - you sense the surface is 0.3 mm high and lower the tool by 0.3 mm. Feedforward predicts and pre-empts - you know this board always cups, so you compensate before cutting. Good systems blend both: predict what you can, measure the rest.

Why does the loop make robots robust? Because it converts a machine that demands a perfect, known world into one that copes with an imperfect, partly-unknown one. An open-loop arm is only as good as its calibration and its stock; a closed-loop arm tolerates warped timber, a mis-set fixture, a block that shifted, a layer that slumped - it notices and corrects. That tolerance to messiness is precisely what a construction site, a reclaimed-material stream, or a natural material like timber, bamboo or stone demands. Feedback is what carries digital fabrication out of the clean lab and into the real, irregular world.

Sensing is not free, and adaptive systems bring their own discipline. Calibration is foundational: a scan is only as trustworthy as the hand-eye calibration relating the sensor to the tool, and a millimetre of calibration error becomes a millimetre of cut error. Latency decides what real-time can do - if the loop reacts slower than the process changes, it lags and can oscillate. Sensors have noise and blind spots: shiny or dark surfaces defeat some scanners, vision needs adequate light, a probe only knows the points it touches. And more feedback means more to tune and more to go wrong; a simple open-loop path can beat a poorly-tuned closed one.

Safety rises with capability, and here it is non-negotiable. An adaptive robot may move in ways a fixed program would not, reacting to what it senses - which makes hard guarding, interlocks and a clear stop essential, and makes never assume the next move the rule around a sensing machine. Collaborative robots with force limiting can share space more safely, but 'collaborative' is a rated capability, not a promise - it depends on speed, payload and tool. Industrial arms carrying spindles, torches or grippers demand fencing, training and supervision, full stop. Adaptive fabrication is powerful precisely because the machine acts on what it perceives; treat that autonomy with the caution it deserves, and leave commissioning and safety sign-off to qualified people.