Why Your Robot Fails in Production: The Demo-to-Deployment Gap Nobody

Talks About The demo floor is a lie. A robot glides across the carpet, identifies a part in a bin, picks it, and places it with millimetre precision. The crowd nods. Investors write checks. Engineers celebrate. Then the machine ships to a warehouse in Dortmund or a factory floor in Detroit, and everything falls apart. This demo-to-deployment gap is not a new problem. It is the central problem in industrial robotics today, and it starts with perception. The lab is not the floor In controlled conditions, robots have every advantage. Lighting is fixed, backgrounds are clean, objects sit at known angles. The perception stack performs beautifully because the world has been arranged to suit it. Real environments offer no such favours. Warehouse floors have shifting sunlight through skylights, reflective stainless steel, transparent packaging, forklift traffic, and vibration from adjacent lines. Each variable exposes a weakness that never appeared in the demo. What looks like a planning failure often begins with a depth map that is confident and wrong. A robot cannot plan around bad data. The sensing layer must be reliable enough to survive the floor, not just the trade show. 2D cameras hit hard limits Traditional 2D vision remains useful for barcode reading, inspection, and object tracking. But a flat image does not measure depth. Depth can be inferred from motion, learned priors, or multi-view geometry, but those estimates collapse when lighting, texture, occlusion, or materials change. A 2D system guessing depth from shadows is not a 3D system. It is a 2D system hoping. This is why 3D vision, depth cameras, and sensor fusion have become central to deployment. Robots need spatial measurements from the physical world, not smarter guesses from flat images. Depth sensing is a toolbox, not a product No single sensor category wins every task. Each technology solves specific problems and introduces specific failure modes. Structured light projects a known pattern and reads deformation to estimate depth. It works well for indoor inspection and measurement. It struggles with ambient light, motion, reflective surfaces, transparent materials, and interference from other active emitters. Stereo vision uses two offset cameras to match corresponding points and compute disparity. Passive stereo depends on texture and light. Active stereo adds infrared projection for low-texture scenes. Both suffer from motion blur, repetitive patterns, occlusion, and range trade-offs. Time-of-flight measures returning infrared light to estimate distance. ToF cameras are compact and fast, but ambient infrared, multipath reflections, and reflective surfaces distort results. Lidar, RGB cameras, and inertial measurement units each have valid roles. The right stack depends on task, range, lighting, materials, motion, compute budget, safety requirements, and what happens when the system fails. AI improves perception. It does not replace measurement. AI can denoise depth maps, fill gaps, fuse RGB and depth, estimate pose, and track motion. These are genuine advances. But AI still depends on reliable physical data. A neural network cannot hallucinate a correct depth estimate from garbage input. The practical implication is that sensor selection and AI processing must be designed together. A better model cannot compensate for a sensor that fails under the actual operating conditions. What engineers should do differently First, test perception under the real floor conditions, not a cleaned-up subset. Second, characterise failure modes explicitly. Know what makes the depth map drop out, and have a recovery behaviour. Third, design the sensor stack for the task, not for the demo. A system that works in 80% of conditions and fails silently in 20% is worse than one that works in 60% and reports uncertainty honestly. The robots that survive deployment are not the ones that looked best on the show floor. They are the ones built by engineers who understood that the floor does not care about the demo.

M4S TAKE

My take: AI claims need scrutiny. The useful implementations reduce cycle time or defect rates in measurable ways. Vague promises about 'optimization' without specific metrics are usually marketing.

Simon McLoughlin