Vision-Guided Deburring: Teaching The Tool Where The Edge Is

Overview

Deburring a casting with a robot has always required knowing where the part is. The traditional answer is to teach the path by hand on a sample part, record the positions, and rely on the fixture to present every subsequent part in the same place. That works when volumes are high and geometry is stable, and it fails when a programme runs many parts at low volume.

A vision-guided cell replaces the taught path with a measured one. It scans the part, computes where it actually sits in three dimensions, and generates a tool path from that measurement rather than from a stored recording. The result is a cell that can be turned to a new part without hours of teaching, at the cost of a scan and a computation in every cycle.

This article describes what that change buys, what it costs in cycle time, and where teaching remains the better approach. Big Bird Industrial (formerly Shimada Big Bird Industrial) has been building cleaning and deburring systems for automotive powertrain components since 2001, and the configuration described here follows the vision-guided deburring systems introduced for large and irregular parts.

1. Why Deburring Is Usually Taught

Teaching exists because a robot has no way of knowing where a burr is unless someone shows it.

The Cost Of A Taught Path

Teaching a path means driving the tool along every edge that needs attention, at every orientation, and recording the result. On a part with dozens of edges that is a shift of work, and it has to be repeated whenever the part, the fixture or the tool changes. For a high-volume part amortised over years, that is negligible. For a part produced in hundreds, it is not.

The Assumption That Breaks

A taught path assumes the part arrives where the fixture says it will. Castings vary: a core shift moves an edge by a millimetre, draft variation moves a wall, and a fixture that holds the part on a rough surface repeats less precisely than one holding it on a machined face. A rigid path then cuts air on one part and digs into the next.

2. What Vision Changes

The measured approach replaces position certainty with position measurement, which changes the failure mode rather than removing it.

Locating The Part In Three Dimensions

A scan of the part produces a representation of its actual surfaces. Comparing that to the nominal model gives a pose: where the part sits and how it is rotated relative to the cell. That pose then transforms every path point, so the tool follows the real edge rather than the drawing's edge.

Generating The Path

With the pose known, the path can be generated from the nominal model rather than recorded from a sample. The engineering effort moves from the shop floor to the office: instead of teaching a robot, an engineer defines which edges need what treatment, and the cell works out the motion.

3. Where The Time Goes

A vision-guided cell runs in the range of five to ten minutes per part on large or irregular components, and that figure is dominated by scanning and by tool motion rather than by computation.

Scanning Is The New Setup

The scan replaces the teaching session, but it happens every cycle rather than once. On a large part that is a substantial fraction of the total time, which is why the scan strategy matters: scanning only the regions that carry edges to be deburred is much faster than scanning the whole component.

When It Pays

The economics favour vision when the teaching time per part exceeds the scanning time per part. A part produced in dozens with dozens of edges is the clear case: teaching would cost more than the production run. A part produced in thousands with a stable fixture is the opposite case, where a taught path costs nothing per part and vision adds time to every cycle.

4. Comparing The Two Approaches

The comparison is not about which method is more advanced. It is about which cost dominates for a given part.

5. What Accuracy The Scan Needs

The scan does not need machining accuracy. It needs enough accuracy to put the tool on the edge, which is a looser requirement than measuring the edge.

Two Different Tolerances

Locating the part requires accuracy comparable to the fixture repeatability it replaces. Following an edge requires accuracy comparable to the tool's own compliance, because a brush or a cutter absorbs some positional error by design. Treating both as the same requirement is what produces over-specified and unnecessarily expensive scan systems.

Why The Tool Tolerance Helps

A compliant tool tolerates a positional error that a rigid one would turn into a gouge. That is why vision-guided deburring is practical with a scan accuracy that would be inadequate for a rigid machining operation: the process absorbs what the measurement misses.

6. Verifying A Guided Cell

Verification has two parts: does the cell find the edge, and does it remove the burr without damaging the surface next to it.

The first is tested by presenting parts with deliberately displaced geometry, simulating a core shift, and confirming that the tool still follows the edge. The second is tested on parts with burrs of known size at known locations, checking both that the burr is gone and that the adjacent surface is unchanged. A cell that passes the first test and fails the second has a path generation problem rather than a vision problem.

7. Where It Still Needs A Person

Vision guidance automates the motion, not the judgement. Deciding which burrs matter, what edge condition is acceptable and which surfaces must not be touched remains a decision made from a drawing and a sample, and it is the input the cell cannot generate for itself.

That judgement is also what determines whether the cell is fast. A cell told to deburr every edge runs longer than one told to deburr the edges that matter, and the difference is a decision rather than a capability.

8. What The Cell Needs From The Part Model

A guided cell generates its path from a nominal model, so the quality of that model becomes a process input rather than a document. A model that is out of date by one revision produces a path along an edge that no longer exists, and the cell will follow it faithfully.

Edge Lists Versus Whole Models

The practical approach is not to feed the cell an entire CAD assembly but a list of edges with a treatment for each: a radius, a chamfer, a brush pass. That list is short, reviewable and version-controlled, which makes it possible to confirm that the cell is deburring what the drawing requires rather than what the model happened to contain.

Keeping the list under revision control also answers the question that arises after a design change: whether the cell was retaught. A record of the list version used at each production period turns that question into a lookup.

9. Cycle And Throughput On A Guided Cell

A cell running five to ten minutes per part is not a high-volume cell, and that is a consequence of the choice rather than a defect in it. The scan and the tool motion together set the cycle, and both scale with the number of edges rather than with the size of the part.

That means the throughput question is answered by the edge list. A part with twelve edges to deburr runs in half the time of one with twenty-four, and the difference is a decision about what the drawing requires rather than a hardware limit. On a cell that serves several variants, the edges per variant become the planning unit rather than the parts per hour.

Where the volume needs to rise, the usual answer is a second cell working on the same programme rather than a faster one, because the scan and the tool motion both resist compression.

Frequently Asked Questions

Can Vision-Guided Deburring Handle Any Part?

Any part that can be scanned from the positions available and whose edges can be reached by the tool. Deep internal edges and edges inside closed sections remain difficult for both methods.

Does It Remove The Need For A Fixture?

It reduces the precision the fixture must provide, because the cell measures where the part is. It does not remove the need to hold the part against tool forces.

How Long Does A Scan Take?

It depends on the area scanned and the sensor. Scanning only the regions that carry edges is the main lever, and on large parts that reduction is what makes the cycle viable.

What Happens If The Scan Is Wrong?

The tool follows a path based on a wrong pose, which produces either a missed edge or a mark on an adjacent surface. That is why the verification includes deliberately displaced parts rather than only nominal ones.

Conclusion

Vision-guided deburring replaces a taught path with a measured one, which moves the cost from setup to cycle time. It wins where the teaching time per part exceeds the scanning time per part, which in practice means low volume, large or irregular components.

The engineering judgement that remains is which edges deserve attention. A cell that deburrs everything is slower and no better than one that deburrs what matters, and that decision comes from the drawing rather than from the sensor.