Latest / The Robotics Podcast with Fexingo: Autonomous Systems, Industrial Robots, and Hardware / Why Your Robot Vacuum Still Bumps Into Things
Transcript
- Lucas: Alright, let's start with something that annoys almost everyone who owns one. You buy a robot vacuum — maybe a Roomba, maybe a Roborock, maybe something cheaper — you set it up, press the button, and within three minutes it's thumping into the leg of your coffee table. Not once. Again and again. Luna: Right, and you think — wait, it's 2026. My phone knows exactly where I am. Why doesn't this thing know where the table is? Lucas: That's the exact question. And the answer is fascinating because it's not about technology being 'behind.' It's about a conscious engineering trade-off that's been in place for over twenty years. The bump and go behavior you see — that little tap, back up, turn — is actually a deliberate feature, not a bug. Luna: A feature? Feels like a bug when I hear it bumping at 2 AM. Lucas: I know. But here's the thing: a simple bumper switch — literally a mechanical microswitch behind a plastic flap — is absurdly reliable. It costs maybe ten cents. It works in total darkness. It works if the sensor is covered in dust. And it tells you, unambiguously, 'I am touching something.' That's information most other sensors can't give you with that level of certainty. Luna: But LIDAR — the spinning laser thing on top of the more expensive vacuums — that's supposed to map the room and avoid collisions entirely, right? Lucas: In theory, yes. In practice, LIDAR has blind spots. Low-profile obstacles — like a charging cable, or a stray sock, or the leg of a chair that's narrow enough — can fall below the laser plane or between scan lines. The bumper catches exactly those things. So even a robot with a full LIDAR rig usually keeps a bumper as a backup. Luna: So you're saying the bump is a safety net. But why can't they just add more sensors to cover the gaps? Lucas: Cost and complexity. iRobot's first Roomba in 2002 used only bumpers and infrared cliff sensors — no cameras, no lasers. It cost two hundred dollars. Compare that to a Waymo car, which has something like five lidar units, twenty cameras, radar — and costs well over a hundred thousand dollars. For a consumer robot at a few hundred bucks, you cannot put a ten thousand dollar sensor suite on it. Luna: So the bumper is the pragmatic compromise. But has anyone actually tried to solve the problem without bumps — like, in a research setting? Lucas: Yes, and some of the work is really clever. There's a 2018 study from Carnegie Mellon where they put tactile sensors — basically artificial skin with pressure sensitivity — on the front of a robot. The robot could detect contact before it fully collided, and distinguish between a hard surface like a wall and a soft one like a person. But the sensors were fragile and expensive. They haven't made it to production. Luna: Right. And even if they did, you'd still have the problem of corners and edges that the skin doesn't cover. Lucas: Exactly. So the bumper remains. And it's not just vacuums. Industrial automated guided vehicles — the ones that move pallets in warehouses — also use bumpers. They're called 'contact safety edges' in that world. They're required by safety standards in many countries. If a human walks in front of an AGV, the bumper stops it instantly. LIDAR can miss a person in certain lighting or if they're wearing dark clothing. The bumper is the last line of defense. Luna: That actually makes a lot of sense. But it still feels like a hack. Why can't we have both — cheap LIDAR and a better bumper? Lucas: We do have both on many robots. The higher-end Roombas and competitors have a LIDAR unit on top AND a bumper. The bumper is just the final layer. The problem is that LIDAR alone can't guarantee zero collisions because the environment changes. You map the room, but then someone puts a chair in a different spot, or the dog leaves a toy. The robot's map is outdated. The bumper catches that. Luna: So the bump tells the robot 'hey, the map is wrong, update it.' That's actually smart. Lucas: Exactly. It's a feedback mechanism. And there's interesting research from MIT — about 2021 — where they used the bumper data itself to improve the map. Every time the robot bumps something, it marks that location as an obstacle and adjusts its future path. Over a few cleaning cycles, the robot learns where the furniture is and bumps less and less. But it never stops entirely because nothing is permanent. Luna: Right. So the bumper is almost like a learning sensor. But it's still annoying to watch. Lucas: It is annoying. And some companies are trying to reduce it with better software. For instance, using a camera with deep learning to recognize common obstacles like cables and shoes. But those models still fail on novel objects. A bumper never fails — it's a physical switch. That reliability is hard to beat. Luna: How much better could it get if we just threw more computing power at the problem? Like, an iPhone has a neural engine now — why not put that on a vacuum? Lucas: You could. And some do. But the power budget is tight. A vacuum runs on a battery, and running a neural network continuously would drain it faster. The bumper uses zero power until it's pressed. So it's also an efficiency win. Luna: Interesting. So the bumper is cheap, reliable, low-power, and provides a clear signal. No wonder it's still around. Lucas: It's a beautiful example of an engineering trade-off where the simplest solution is still the best for the use case. And it's not just vacuums. You see the same logic in robot lawn mowers, pool cleaners, even some warehouse robots. They all bump. It's not a failure of technology — it's a deliberate design choice that prioritizes cost and reliability over perfect gracefulness. Luna: I guess I'll stop yelling at my Roomba then. But what about the future? Could we ever get to zero-bump robots? Lucas: Maybe. But it would require a sensor that is as cheap, as reliable, and as unambiguous as the bumper. I'm not sure that exists yet. There's work on time of flight depth sensors at very low cost — the kind used in smartphone face unlock — but they have limited range and can be confused by transparent surfaces. A bumper doesn't care if something is glass or wood. Luna: Right. So we're stuck with bumpers for a while. But at least now I understand why. Lucas: And understanding the 'why' is the whole point. By the way, if this kind of deep dive into how technology really works is valuable to you — something you'd share with a friend or think about next time you see a robot — we'd love to keep making these. A small group of listeners already helps make that possible through buy me a coffee dot com slash fexingo. No pressure, just if you find value in the show and want it to stay ad-free and independent. Luna: Yeah, it's a really small gesture that goes a long way. We put a lot of research into each episode, and knowing it helps people understand the tech they use every day makes it worth it. Lucas: Alright, back to the bumps. One last thing — there's actually a funny consequence of bumper-based navigation. Because the robot has to bump to learn, the first few cleaning cycles are almost comically inefficient. But over time, it gets faster. iRobot published data showing that after ten cleaning cycles, a Roomba covers the same area in about half the time as the first run. The bumps teach it. Luna: So it learns from its mistakes. That's kind of human. Lucas: It is. And it's a reminder that in robotics, sometimes the most elegant solution is the one that accepts imperfection and learns from it. Not every problem needs a laser.