Latest / The Robotics Podcast with Fexingo: Autonomous Systems, Industrial Robots, and Hardware / Why Robot Sensors Fail in High Humidity
Transcript
- Lucas: There's a seafood processing plant in Gloucester, Massachusetts, that installed a robotic vision-guided gripper system in early 2025 to pack cod fillets. By April, the system was dropping twelve percent of its catch. Not a software bug, not a calibration issue — humidity. Luna: Condensation on the sensor lens. I've heard about this in cold storage warehouses too, where the temperature differential fogs things up. Lucas: Exactly. The plant's processing floor runs at about four degrees Celsius, high humidity from the wash-down cycles, and the robot's camera housing warms up just enough from its own electronics that when it moves into the cold zone, water vapor condenses on the glass. The vision system sees blur, misidentifies fillet edges, and the gripper misses the grip point. Luna: So it's a physics problem, not a software problem. The sensor fusion stack could be perfect, but if the input's fogged, garbage in, garbage out. Lucas: Right. And this is surprisingly common. A 2023 survey from the Robotic Industries Association found that about fifteen percent of unplanned downtime in food and beverage robotics was traced to sensor degradation from moisture or condensation. It's not a headline-grabbing failure mode, but it eats into throughput. Luna: What about LiDAR? Does it have the same problem? I know some outdoor autonomous vehicles use spinning LiDAR that's sealed. Lucas: LiDAR actually has a related issue. The spinning housings often have vents for pressure equalization, and in high humidity, moisture can seep in and frost onto the internal optics. A 2024 paper from MIT's Computer Science and Artificial Intelligence Lab tested several commercial LiDAR units at ninety-five percent relative humidity and found that after thirty minutes, range accuracy dropped by up to eight percent. The beam scatters off water droplets on the emitter window. Luna: Eight percent might not sound catastrophic, but for a robot navigating tight warehouse aisles or performing precision pick and place, that's enough to cause collisions or missed picks. Lucas: Exactly. And the mitigation strategies are still fairly crude. Some manufacturers use hydrophobic coatings — basically a nano-scale water-repellent layer on the lens. But those wear off with cleaning cycles. Others use heated housings that keep the glass above the dew point, but that draws power and adds thermal management complexity. Luna: I remember reading about a startup that tried ultrasonic vibration to shake droplets off the lens. Did that go anywhere? Lucas: It did, actually. A company called DropletClear — they were acquired by a larger sensor supplier in 2022. Their approach used a piezoelectric element bonded to the lens cover that vibrated at around fifty kilohertz, shedding water droplets. It worked well in lab tests, but in field conditions with greasy residues from food processing, the vibration wasn't enough to break the surface tension. So it's still niche. Luna: So we're basically stuck with wiping lenses manually, or designing robots that avoid the humid zones altogether. Lucas: Which is what a lot of facilities end up doing. They reposition the sensor further from the source of steam or spray, add an air knife that blows a continuous stream of dry air across the lens. That works, but it adds cost and complexity. And it doesn't solve the problem for mobile robots that have to go everywhere in the facility. Luna: This reminds me of something we've touched on before — hardware reliability is often the bottleneck that software people forget about. A neural network can be state of the art, but if the camera's fogged, it's blind. Lucas: Yeah. And the thing is, this problem extends beyond food processing. Agricultural robots working in early morning dew, offshore inspection drones in salt spray, even warehouse robots in unconditioned spaces near loading docks — they all face some version of this. The sensor fusion community talks a lot about algorithm robustness, but the physical robustness of the sensor itself is just as critical. Luna: I want to come back to that agricultural angle in a moment. But first — and this ties in directly — we keep this show ad-free because we think that kind of honest, specific engineering conversation is valuable on its own. No commercials, no sponsors pushing their hardware. If that matters to you, the simplest way to support it is buy me a coffee dot com slash fexingo. That's it. Lucas: And it genuinely makes a difference. Those contributions cover hosting, research, and keeping the podcast independent. No pressure, just the link if you feel like it. Luna: Okay, back to the farm. Lucas, you mentioned agricultural robots dealing with dew. How bad is it there? Lucas: It's a major operational headache. A strawberry-picking robot from a company called Harvest Automation — they operate in California's Central Valley — found that on foggy mornings, the stereo cameras used for depth perception would fog up within minutes. They had to delay harvests until the sun burned off the moisture, which defeated the purpose of having a robot that could work twenty-four seven. Luna: Did they ever solve it? Or is it still a morning-only operation? Lucas: They implemented a heated lens cover that draws about fifteen watts, which keeps the optics clear but adds a non-trivial battery drain. On a robot that already has limited runtime, that's a real trade-off. They also added a dew sensor that tells the robot to wait until conditions improve — basically an automated weather hold. Luna: So we have all this investment in autonomy, and the robot is still sitting idle because of water on a lens. Lucas: Right. And the irony is that the human picker working next to the robot just wipes their glasses on their sleeve and keeps going. That's a hundred million years of evolution giving us a self-cleaning hydrophobic lens — the human eye. Replicating that in a machine is deceptively hard. Luna: There's research into bio-inspired surfaces, isn't there? Lotus leaf effect, that kind of thing. Lucas: There is. Superhydrophobic surfaces with micro- and nano-scale textures that cause water to bead up and roll off. A group at Stanford published a paper in 2024 showing a coating that reduced condensation by sixty percent on a camera window. But the durability issue remains — those textures are fragile and degrade with abrasion. In a food plant where you're cleaning with high-pressure hoses, they don't last long. Luna: So the problem isn't a lack of ideas. It's that every solution has a trade-off: cost, durability, power, complexity. Lucas: Exactly. And that's why this remains a topic of active research and engineering. It's not a solved problem. For any roboticist deploying in an uncontrolled environment, humidity is a first-order constraint. You have to budget for sensor maintenance, or accept some downtime, or design around it. Luna: It also affects indoor robots more than people think. I visited a beverage bottling plant last year — not a robotics facility, but they had AGVs moving pallets. The AGVs used laser scanners for safety, and the scanner windows had to be wiped every shift because of condensation from the bottle-washing station. Lucas: That's a great example. And those safety scanners are critical — if they fail, the robot stops. So you end up with a process where the cleaning schedule becomes part of the robot's uptime calculation. Some facilities now plan for a ten-minute window every four hours for lens cleaning. That's lost productivity. Luna: Is there any industry standard for testing sensor performance under humidity? I know there are IP ratings for ingress protection, but that's about dust and water jets, not condensation. Lucas: Not really. The IP rating tells you whether water can get inside the housing, but it doesn't test for internal fogging or condensation on the optical element. There's an emerging effort — the IEEE is working on a standard for optical sensor reliability in humid environments, but it's still in draft. So right now, buyers have to do their own field testing. Luna: Which is expensive and time-consuming, especially for smaller robotics startups. Lucas: Exactly. And that's a barrier to adoption. If a small manufacturer wants to automate a humid process — say, cheese aging or mushroom growing — they may not have the resources to test five different sensor configurations. They might just give up and stick with manual labor. Luna: So solving this humidity problem isn't just a technical challenge. It's an economic one that affects how quickly robotics can penetrate certain industries. Lucas: I think that's exactly right. The industries that are hardest to automate — food, agriculture, textiles — often have the most challenging environments. Temperature swings, moisture, dust. If we can make sensors that work reliably in those conditions, we unlock a huge amount of automation potential. Luna: And we're not talking about exotic new sensor modalities. Just making the existing ones robust enough to handle a foggy morning. Lucas: Exactly. It's boring engineering. But boring engineering is what makes robots actually useful in the real world. Luna: Alright, so where should a listener go if they want to dig deeper? Any papers or companies to follow? Lucas: The MIT paper I mentioned — 'Effect of Humidity on LiDAR Performance' — is publicly available on their CSAIL site. And for coatings, look up the work from the Aizenberg Lab at Harvard on omniphobic surfaces. They've done some fascinating work on surfaces that repel both water and oil, which is relevant for food processing where greasy residues are an issue. Luna: I'll add one: there's a company called ClearView Optics that makes heated camera housings specifically for robotics. They're not cheap — around five hundred dollars per unit — but they're a drop-in solution for some common camera models. Lucas: Good tip. And I think that's where we'll leave it. The humidity problem isn't solved, but it's being worked on. And the next time you see a robot in a cold aisle or a foggy field, you'll know one of the hidden challenges it's up against. Luna: Thanks for listening. We'll be back next time with another deep dive into the hardware that makes autonomy possible.