Latest / The Robotics Podcast with Fexingo: Autonomous Systems, Industrial Robots, and Hardware / Why Robot Arms Still Cant Handle a Light Switch
Transcript
- Lucas: Luna, I want you to imagine walking into a dark room in your house. You reach out, your fingers find the light switch, and you flip it. That entire sequence takes you maybe half a second. Luna: Yeah, it's so automatic I don't even think about it. Lucas: Right. Now imagine a robot doing the same thing. A state of the art robotic arm, with cameras, force sensors, and a gripper designed to handle delicate objects. A 2025 study out of MIT's CSAIL put that exact scenario to the test. The robot succeeded only 23 percent of the time. Luna: Twenty-three percent? That's abysmal. What went wrong? Lucas: Almost everything. The study used a standard KUKA LBR iiwa — that's a seven-axis collaborative robot arm with pretty good torque sensors. They programmed it to locate a wall plate, align its gripper with the toggle switch, and then flip it. The failures clustered into three categories: misalignment, excessive force, and trajectory errors. Luna: Misalignment I get — the camera might not see the switch precisely. But excessive force? You'd think a robot with torque sensors would know how hard to push. Lucas: That's the paradox. The torque sensors on the iiwa are designed for low-force tasks like assembly or polishing. But a light switch requires a very specific force profile: you need enough force to overcome the detent — that little click mechanism — but not so much that you snap the switch or bend the wall plate. The robot would often either push too gently and not flip it, or push so hard it broke the switch. Luna: So it's a Goldilocks problem. And trajectory errors? Lucas: The robot's arm moves in a straight line from point A to point B. But human fingers naturally follow a curved arc when flipping a switch — we rotate our wrist slightly as we push. That arc reduces the lateral force on the switch. The robot's linear motion applies a shearing force that can jam the toggle. The MIT team tried adding a curved trajectory, but the robot couldn't adapt to different switch orientations on the fly. Luna: And this is a KUKA iiwa — one of the most advanced collaborative robots on the market. What about cheaper robots? Lucas: They didn't test those, but you can imagine. The iiwa has built-in compliance — it can back off when it senses resistance. Most industrial arms are stiff position-controlled machines. They'd either fail to move the switch or punch a hole in the drywall. The study essentially showed that even with the best hardware, the control software isn't general enough to handle the variability of a simple light switch. Luna: Variability meaning different types of switches? Toggle versus rocker versus dimmer? Lucas: Exactly. In the US alone, there are dozens of standard switch designs. The MIT team tested six common types. The robot managed decently on a large toggle switch — about a 40 percent success rate. But on a small rocker switch, it dropped to 12 percent. And dimmer switches with sliders? Zero percent. The robot couldn't even locate the slider consistently. Luna: So the problem isn't just force control — it's perception. The robot struggles to identify the switch type and adjust its strategy. Lucas: Right. And that's where deep learning comes in. The MIT team trained a neural network on thousands of images of switches in varied lighting and angles. But when they deployed it on the robot, the model couldn't generalise to switches it hadn't seen. It would mistake a light switch for a thermostat or an outlet. So the robot would try to 'flip' a wall outlet — which, of course, doesn't move. Luna: That's almost funny. But it highlights a core issue: robots are terrible at handling objects that don't have a consistent shape and behavior. A switch in a home might be at a different height, surrounded by different materials, painted over — all things a human ignores effortlessly. Lucas: Yes. And that's the deeper takeaway from this episode. This isn't just about light switches. It's about why the vision of a general-purpose home robot — something that can do dishes, fold laundry, flip switches — remains so distant. Each of those tasks has its own set of physical and perceptual challenges. The switch problem is a microcosm. Luna: And the industrial robots we have today are incredibly good at doing one thing in a controlled environment. But the home is chaos. Lucas: Exactly. An automotive assembly line robot can weld a car body with micron precision because every part is exactly where it's supposed to be. But put that same robot in a kitchen and ask it to flip a switch? It's lost. The MIT study is a useful reality check for the robotics field. Luna: So what's the path forward? How do we get from 23 percent to, say, 90 percent? Lucas: The researchers suggest two main avenues. First, better sensing — not just vision, but tactile sensing. If the robot can feel the switch's orientation and resistance in real time, it can adapt its force and trajectory on the fly. Second, more compliant hardware — arms that can actively vary their stiffness, like a human arm can. There's work at Stanford on variable impedance actuators that could help. Luna: And on the software side? Deep learning that can generalise from fewer examples? Lucas: That's the holy grail. The MIT team tried a technique called domain randomization — training the network on synthetic images with random backgrounds, lighting, and switch colors. It helped a bit, but not enough. The robot's success rate went from 23 to 31 percent. Still far from usable. Luna: So we're still years away from a robot that can turn on the lights for you. Lucas: Probably. But it's worth noting that some companies are already deploying robots in semi-structured environments like hotels and hospitals. They use docks and predefined paths. A light switch in a hotel room is always in the same spot — the robot can be pre-programmed. The real challenge is the unstructured home. Luna: Right. And that's where the market is. Everyone wants a Rosie the Robot, but we're stuck with Roombas. Lucas: That's a good comparison. The Roomba succeeds because it doesn't have to manipulate anything — it just vacuums. The moment you add manipulation, you add an order of magnitude more complexity. Light switches are just the tip of the iceberg. Luna: And speaking of the show, if today's deep dive into robotics gave you something to think about — maybe a new appreciation for how hard it is to flip a switch — we want you to know that listener support is what keeps this podcast ad-free and focused on topics like this. It's a small thing, but it helps us keep going. Lucas: Yeah, absolutely. If you find value in these conversations, you can support the show at buy me a coffee dot com slash fexingo. No pressure at all — just a way for those who want to help. Luna: And we really appreciate it. Now, back to the switch problem — one more angle. The MIT team also tried using a compliant gripper made of soft silicone. Did that help? Lucas: It did, actually. The soft gripper could conform to the switch shape and absorb some of the misalignment. Success rates went up to about 45 percent. But the gripper lacked the rigidity to apply enough force for some switches. So there's a trade-off between compliance and strength. Luna: Sounds like we need a gripper that can modulate its stiffness in real time. Lucas: Exactly. Researchers are working on that — it's called variable stiffness grippers. Some use jamming of granular materials, others use shape-memory alloys. But none are commercially mature yet. Luna: So for now, if you want a robot to turn on the lights, you might be better off using a smart bulb and a voice assistant. Lucas: That's the irony. We've solved the problem of controlling lights without a physical switch, but we can't get a robot to operate the switch itself. It's a reminder that sometimes the most 'simple' human actions are the hardest to replicate. Luna: Thanks for listening to The Robotics Podcast with Fexingo. We'll be back next week with another episode.