GERMANY VIDEO REVIEW · VISION LOADING
One Camera Sweep and It Grips: RoboJob's AI Vision Loading, Live-Tested at EMO
01What this live-test video is about
Voice of CNC's second RoboJob video in its EMO 2025 series stars the new Pallet Load Integrate system—an automation solution where a robot grabs workpieces from a material pallet and loads them onto the machine. Unlike a brochure-style intro, host Alex tests it live (testet live) on the floor, and even adds a "workpiece-shift challenge (Bauteil-Verschiebe-Challenge)": shifting the workpieces on the pallet to see whether the system can still find and grab them[1].
The capability list in the description reads like a spec sheet[1]: a 2D camera generates 3D spatial data; AI recognizes workpieces on a wooden pallet, with reflections and texture not affecting it; no pre-lining-up, no sorting—push the pallet in and it starts; setup in under 5 minutes; suitable for forgings, heavy parts, and varying geometry.
02Technical breakdown: 2D camera, 3D space, AI recognition
Every item on the list corresponds to a concrete technical choice:
- 2D camera → 3D spatial data: instead of an expensive 3D scanner, it reconstructs spatial information from an ordinary camera's images—a friendly cost structure, and the hardware prerequisite for "5-minute setup"[1].
- AI workpiece recognition: a deep-learning vision model recognizes workpiece contours and pose from images—"reflections and texture don't matter" is precisely deep learning's advantage over traditional machine vision (which relies on stable lighting and contrast)[3]. The wooden-pallet detail is very honest: that's exactly how material sits in a real factory.
- No lining-up, no teaching: because the system "understands" for itself what's on the pallet and where, the two traditional prerequisites of "arrange the material neatly for the robot" and "teach it batch by batch" both disappear at once[4].
Theoretically, this is the same family of technology as AI drawing reading—feature-recognition research (such as FeatureNet) has proven that neural networks can "understand geometry" with high accuracy[2]; the only difference is whether the input is a camera image or an engineering drawing. The machine's ability to understand geometry is rewriting both the "machine-side" and "off-machine" stages of the flow at the same time.
03Which costs it dismantles
The hidden-cost list of traditional robot load/unload is exactly why high-mix small-batch shops shy away from automation:
| Traditional prerequisite | Hidden cost | After vision AI |
|---|---|---|
| Workpieces need lining-up and positioning | Lining-up fixtures, racks, manual placement | Push the pallet straight in, the AI finds them[1] |
| Batch-by-batch teaching | Engineer time per batch change | Recognition replaces teaching, setup in minutes[1] |
| Amortization over batch size | Small batches can't spread the upfront cost | The upfront cost itself is compressed; small batches start to pay off |
This is exactly the process by which the "what parts aren't suited to automation" list in Robot loading/unloading primer gets rewritten—the boundary is moving, and the criteria need to update with it.
04The evaluation angle for Taiwan's shops
After watching this test, when a Taiwanese high-mix small-batch shop evaluates load/unload automation, its list of questions should update to:
- How much does a batch change cost? Don't just ask about cycle time—ask "what does a person do for a batch change, and how long does it take"—this is where the vision solution's core value lies.
- Can it handle your material situation? Take your messiest pallet (mixed material, reflective parts, stacking) and try it—the video's "shift challenge" is exactly the right testing mindset.
- Can program supply keep up? The old question once more: once the robot loads the material onto the machine, the machine needs a program to run. Load/unload automation makes the machines churn faster, and the "waiting for programs" bottleneck only becomes more obvious—preparation-chain automation (AI reads drawings, generates code within the in-house tool library and machine constraints, simulates, verifies dimensions) should go in sync or ahead (see The complete guide to CNC auto-programming).
Machine-side vision AI and off-machine drawing-reading AI are essentially two ends of the same thing: let the machine understand geometry, and let people do less translation. Only when both ends go in together does full-flow automation for high-mix small-batch truly take shape.
05FAQ
What's the difference between vision-based loading and traditional robot loading?
Traditional solutions require workpieces to be lined up and positioned (fixtures, racks, manual placement) and taught batch by batch; the vision solution uses a camera plus AI to let the robot recognize the workpieces and their pose on the pallet itself—no lining-up needed, no re-teaching per batch, and the setup time in the video's live test is under 5 minutes. What gets dismantled is exactly the upfront cost that small batches can't spread out.
Is "a 2D camera generating 3D spatial data" reliable?
This is a mature route that reconstructs depth information from 2D images by algorithm, at a cost far below a 3D scanner. Reliability depends on the actual material situation—reflections, texture, stacking. The right way to evaluate it is exactly what the video demonstrates: live-test it with your messiest pallet, and even deliberately move a workpiece to see how the system reacts.
Load/unload automation or program-preparation automation—which comes first?
If your machines are often waiting for programs, do the preparation chain first—load/unload automation makes the machines churn faster, and the waiting-for-programs bottleneck will only get bigger. The ideal is both ends in sync: machine-side vision AI handles the variation in material, off-machine drawing-reading AI handles the variation in drawings, and both are essentially "letting the machine understand geometry"—together they make up complete high-mix automation.
Get notified when new articles and video reviews go live—no inbox flooding, one-click unsubscribe. (newsletter in Chinese)
GEOMETRY UNDERSTOOD
Machine-side, let AI recognize the workpiece; off-machine, let AI read the drawing
Before load/unload automation, make sure program supply keeps up: AI reads drawings, generates code, simulates, and verifies—use your drawing to measure the preparation chain's speed.
Contact an adoption consultant Training courses06References
- IndustryArena / Voice of CNC (YouTube). RoboJob auf der EMO 2025 - Part2: Palettenhandling next level mit KI & Kamera! (published 2025-11-06, live-tested at EMO 2025). youtube.com/watch?v=en1IlMc_5io
- Zhang, Z., Jaiswal, P., & Rai, R. (2018). FeatureNet: Machining feature recognition based on 3D Convolutional Neural Network. Computer-Aided Design, 101, 12–22.
- Wang, J., Ma, Y., Zhang, L., Gao, R. X., & Wu, D. (2018). Deep learning for smart manufacturing: Methods and applications. Journal of Manufacturing Systems, 48, 144–156.
- Groover, M. P. (2019). Automation, Production Systems, and Computer-Integrated Manufacturing (5th ed.). Pearson.