Robot Software
Offline Programming vs Teaching: Where the Break-Even Sits
Teaching wins below roughly 40 waypoints, offline wins above 200, and calibration is what decides whether the offline path lands on the part at all.
Below about 40 waypoints, teaching on the pendant is faster than setting up an offline project. Above roughly 200 waypoints, or whenever the geometry comes from CAD, offline wins decisively. The variable that decides whether offline works at all is not the software but cell calibration, which typically takes half a day and turns a 3 mm mismatch into a 0.3 mm one.
The break-even arithmetic
| Waypoints | Teaching | Offline setup | Offline touch-up | Winner |
|---|---|---|---|---|
| 10 | 0.5 h | 4 h | 0.3 h | Teaching |
| 40 | 2.5 h | 5 h | 0.7 h | Teaching, narrowly |
| 100 | 7 h | 6 h | 1.2 h | Roughly level |
| 200 | 15 h | 7 h | 2 h | Offline |
| 800 | 65 h | 12 h | 4 h | Offline, decisively |
| Regenerated for a revision | as above, again | 0.5 h | 1 h | Offline, overwhelmingly |
The last row is the one that changes procurement decisions. Teaching produces a program; offline produces a program generator. When the part revision changes, the taught program is redone from scratch while the offline path is regenerated in under an hour.
Calibration is the whole game
An offline path is a set of absolute coordinates. Executing them correctly requires the virtual cell and the physical cell to agree, and three sources of disagreement have to be removed in order.
- Tool centre point. A four or six point procedure on the pendant, 10 to 30 minutes. This is the largest single contributor on most cells, because the tool frame is usually estimated from a drawing rather than measured.
- Workpiece and fixture frames. Tying the model geometry to where the fixture actually sits, 20 to 60 minutes. It also makes the program survive a fixture being moved.
- Robot kinematic error. Link length tolerances, joint offsets and deflection. Measuring the arm across dozens of poses with a tracker and fitting corrected parameters takes roughly half a day and brings absolute accuracy from 0.5 to 3 mm down to 0.1 to 0.3 mm.
Skipping step three is legitimate for tasks with wide tolerance, such as painting or coarse deburring. It is not legitimate for a weld seam or an insertion, and that is where most disappointed offline projects went wrong.
The hybrid that most cells actually use
Pure offline and pure teaching are both rare in production. The working pattern splits the path by tolerance.
| Path segment | Tolerance | Method | Share of points |
|---|---|---|---|
| Approach and transit moves | 5 to 20 mm | Offline, never touched | 50 to 70 % |
| Process path, seam or contour | 0.3 to 1 mm | Offline plus calibration | 20 to 40 % |
| Pick and insertion points | 0.05 to 0.3 mm | Taught or probed on the real cell | 5 to 15 % |
Structuring a program this way means a fixture move requires re-teaching only the 5 % to 15 % of points that are tolerance-critical, while the rest regenerate. Programs written as one flat list of taught points offer no such option, which is why structure matters more than method.
When offline is the only option
- The cell does not exist yet. Reach, collision and singularity problems discovered after installation are expensive; discovered in a model they are free.
- The path follows a CAD contour. Weld seams, glue beads, deburring edges and inspection paths are defined by geometry, not by points a human would pick.
- The program must run on more than one robot. Taught points do not transfer between arms without calibration, because each has different systematic error.
- Cycle time has to be quoted before purchase. Vendor offline tools model the real controller motion, which is the only credible source for a cycle estimate.
- The part changes often. High-mix production is precisely the case where regenerating beats reteaching.
Frequently asked questions
When is offline programming worth it?
Above roughly 200 waypoints, whenever the path comes from CAD geometry, when the cell does not exist yet, or when the program will be regenerated for part revisions. Below about 40 waypoints, teaching on the pendant is faster.
Why do offline paths miss the part?
Because the virtual and physical cells disagree. Tool centre point error, fixture frame error and robot kinematic error together produce 0.5 mm to 3 mm of mismatch on an uncalibrated cell. Full calibration reduces that to roughly 0.1 mm to 0.3 mm.
How long does cell calibration take?
Tool centre point calibration is 10 to 30 minutes, fixture frames another 20 to 60 minutes, and full kinematic calibration with a measurement system about half a day. The first two are always worth doing.
Can I transfer a taught program to an identical robot?
Not reliably. Two arms of the same model have different systematic error, so taught points arrive with an offset. Kinematic calibration on both machines is what makes program portability practical.
Should cycle time come from an offline tool or a physics simulator?
From the vendor offline tool, which models the actual controller trajectory generator, blending and acceleration limits. A general physics simulator animates the motion without reproducing the real motion planner.
Sources
- ISO 9283, manipulating industrial robots, performance criteria and test methodsPose accuracy and repeatability definitions underlying the calibration discussion
- Robotics at NISTNational Institute of Standards and Technology, robot performance measurement and test methods
- ISO 8373, robotics vocabularyInternational Organization for Standardization, definitions of tool centre point and frames