Latest / The Robotics Podcast with Fexingo: Autonomous Systems, Industrial Robots, and Hardware / Why Robot Arms Still Cant Handle a Screwdriver
Transcript
- Lucas: You know that moment when you're assembling furniture, and you reach for a screwdriver, and it just works? Your hand aligns the tool, applies just the right torque, and you don't even think about it. Luna: Right, I do that all the time. But for a robot arm — that's apparently a nightmare. Lucas: Exactly. Today we're talking about why robot arms still can't handle a screwdriver. This is episode 78 in our ongoing series on the limits of robotic manipulation, and screwdrivers are a perfect example of a deceptively hard problem. Luna: What makes it so difficult? I mean, we have robots that can assemble car engines, right? Lucas: We do, but those are highly controlled operations — the parts are precisely positioned, the tools are custom, and the screwdrivers are usually automated screwdrivers, not the handheld kind. The problem we're talking about is a general-purpose robot arm picking up a standard Phillips-head screwdriver and driving a screw into wood or metal. That involves force control, alignment, and tactile feedback that current systems just don't have. Luna: So it's not just about gripping the screwdriver — it's about using it correctly. Lucas: Right. There's a 2025 paper from MIT's CSAIL lab — led by researcher Yifan Zhang — that tested a KUKA LBR iiwa robot arm on 50 different screw types. They had the robot pick up a screwdriver, align it with a screw, and drive it in. The success rate across all trials was 73 percent. That sounds decent, but for industrial applications, you need 99.99 percent. And the failures were instructive. Luna: What were the main failure modes? Lucas: Three big ones: cam-out, cross-threading, and bit slippage. Cam-out is when the screwdriver bit slips out of the screw head under torque — common with Phillips heads. Cross-threading is when the screw goes in at a slight angle, damaging the threads. Bit slippage is when the driver bit rotates inside the screw head without engaging. All three come down to force control and alignment. Luna: So the robot doesn't know how much force to apply in which direction. Lucas: Exactly. Humans have an intuitive sense of 'push down while turning' — we adjust in real time based on feel. The robot has torque sensors, but those only measure rotational force. They don't measure lateral forces or the exact position of the bit relative to the screw head. The MIT team developed a new algorithm called ScrewNet that combines force-torque sensing with visual feedback to estimate the alignment state. Luna: How does ScrewNet work? Lucas: It uses a deep neural network trained on thousands of screwdriving attempts. The robot first does a visual scan of the screw head, then makes an initial approach. During insertion, it monitors the force-torque signature — a specific pattern for successful driving versus cam-out or cross-threading. If it detects a failure pattern, it can back out and retry with a corrected angle. The paper reports that ScrewNet improved success rates from 73 percent to 89 percent. Luna: That's a big jump, but still not 99.9 percent. Lucas: No, and the remaining failures are tough. Some are due to variability in screw heads — a slightly worn Phillips head, or a screw that's been painted over. Others are due to material differences — driving into hardwood versus softwood requires different torque profiles. The robot can't sense the material before starting. Luna: So it's a sensing problem primarily. Lucas: Partially. But also an actuation problem. The KUKA arm they used has seven degrees of freedom and very precise torque control. But it lacks tactile feedback — it doesn't have sensors in the gripper that can feel the texture of the screw head or the resistance of the material. There's work on tactile skins for robot fingers, but they're not robust enough for dirty or oily environments. Luna: And this matters for more than just furniture assembly. Lucas: Absolutely. Think about electronics manufacturing — screwing tiny screws into smartphones. Or aerospace — where every bolt has a precise torque specification. Currently, those tasks are done by humans with torque wrenches, or by specialized screwdriving robots that are custom-built for one screw type. A general-purpose robot that can handle any screwdriver would be a game-changer for flexible manufacturing. Luna: Let me ask — from a listener perspective, this reminds me that a lot of listeners have been asking about how we keep making these deep-dive episodes without a paywall or ads. And honestly, it's because a small group of listeners chips in monthly through buy me a coffee dot com slash fexingo. That's literally what funds the research time and production for shows like this. Lucas: Yeah, and we don't plan to change that. No ads, no sponsors — just listener support. It keeps the conversations honest. Luna: And if you've gotten something useful out of this screwdriver deep dive, that's the reason. Back to the topic — so what's the next step for ScrewNet? Lucas: The MIT team is working on incorporating tactile sensing. They're testing a new gripper with embedded capacitive sensors that can measure shear forces — that's the sideways force when the bit is about to cam out. Early results show that with tactile feedback, success rates on the same 50 screws hit 94 percent. But the sensors add cost and complexity. Luna: So we're still a few years away from a robot that can assemble IKEA furniture? Lucas: Probably more than a few. The IKEA assembly challenge involves not just screwdriving, but also aligning dowels, inserting pegs, and handling panels. There's a 2023 paper from Nanyang Technological University that had a robot assemble an IKEA chair — but it took 20 minutes and the parts were pre-positioned. Still, progress is steady. Luna: What about deep reinforcement learning? Could we just train a robot with simulated screwdriving for a million episodes? Lucas: That's being tried. The problem is the simulation to reality gap. In simulation, you can perfectly model screw heads and torque. In reality, there's dust, grease, worn bits, and material variations. The robot learns behaviors that don't transfer. Some groups are using domain randomization — varying the simulation parameters — but it's still an open challenge. Luna: So the humble screwdriver remains a benchmark for robotic dexterity. Lucas: Exactly. And it's a good one because it tests so many capabilities: grip, alignment, force control, tactile feedback, and error recovery. If we solve the screwdriver, we've solved a lot of other manipulation problems too. Luna: What's a concrete takeaway for our listeners? If they're in robotics or manufacturing, what should they watch for? Lucas: Watch for commercial tactile sensors that are robust and affordable. Right now, the best tactile sensors cost thousands of dollars and are lab-grade. If someone brings a $100 tactile skin to market, it could accelerate general-purpose manipulation by years. Also, keep an eye on ScrewNet — the MIT team plans to open-source the code later this year. Luna: That would be huge for small robotics labs. Lucas: It would. And maybe a few years from now, we'll have a robot that can help us assemble that bookshelf from the flat-pack. Luna: I'll be first in line. Lucas: Same here.