Latest / The Robotics Podcast with Fexingo: Autonomous Systems, Industrial Robots, and Hardware / Why Robot Arms Can't Solder a Circuit Board
Transcript
- Lucas: You'd think soldering would be the perfect job for a robot arm. Precise, repetitive, high heat — exactly the kind of task we offload to machines. And yet, walk into any electronics contract manufacturer today, and you'll still see rows of people with soldering irons doing rework on boards the pick and place machines couldn't finish. Luna: Really? I always assumed all those little surface-mount components were placed and soldered by robots. Lucas: The initial placement and reflow oven step — that's fully automated. But for through-hole components, odd-shaped parts, or any board that needs a last-minute engineering change, humans still do the soldering. And it's not just that companies are slow to automate. There's a fundamental mismatch between what a robot can sense and what soldering demands. Luna: So what's the gap? Is it vision, force, something else? Lucas: It's thermal feedback, mostly. When a human solders, we're constantly adjusting the iron's angle, dwell time, and solder feed rate based on how the joint looks and feels. The molten solder has to wet both the component lead and the pad simultaneously. If the pad is a little colder because of a ground plane — that's a massive heat sink — you need to hold the iron longer. But hold too long and you lift the pad or damage the component. Luna: So the robot would need to feel the temperature of the joint in real time and adapt. Lucas: Exactly. And most industrial robot arms have force-torque sensors at the wrist, but those are measuring the force applied, not the thermal state of the workpiece. You can mount a thermal camera, but by the time the camera sees the temperature, the joint is already cooling. The feedback loop is too slow. Luna: There was a paper from MIT's CSAIL last year that tried to teach a robot to solder a through-hole capacitor. I remember reading about it. Lucas: Right — that was the one where they used a custom gripper with a built-in thermocouple and a vision system trained on thousands of example joints. In the lab, it achieved about an 85 percent success rate. But when they moved it to a production line running at 40 parts per minute, the success rate dropped below 50 percent. The variability in board thickness, solder mask color, and component lead length was just too high. Luna: And 50 percent is useless in production. You need 99.9 percent or better. Lucas: Especially for something like medical devices or aerospace electronics, where a cold joint can cause a failure years later. That's the irony — soldering looks like a simple mechanical task, but it's actually a thermal process that requires continuous sensory integration. The robot can't just follow a trajectory. It has to modulate its behavior based on the physics of the moment. Luna: It reminds me of that earlier episode where we talked about why robots can't fold a fitted sheet. The sheet's shape changes every time you touch it. Soldering is similar — the thermal properties of the board change based on what's already been soldered nearby. Lucas: That's a great connection. And it gets to the broader concept of 'assembly tolerance' — the gap between a structured environment where everything is perfectly repeatable and the real world where parts have dimensional variation, oxidation, flux residue. Humans don't think about it because we just adapt. But for a robot, every board is a slightly different puzzle. Luna: Speaking of puzzles — and this is a bit of a left turn — but do you ever think about how much of what we talk about on this show comes down to the fact that robots lack a certain kind of common sense? They can't generalize from one slightly different instance to the next. Lucas: Absolutely. And that's actually why listener support matters for a show like this. We don't run ads, we don't have a corporate sponsor, and that lets us dig into these specific engineering constraints without having to dumb it down or sell a product. The entire show is supported by listeners who find value in understanding why robotics is still hard. Luna: Yeah, it's a niche but curious audience. And if even a few people chip in, it keeps the lights on. Lucas: So if today's conversation gave you something usable — maybe it's the concept of thermal feedback, or the MIT soldering study — and you want to help us keep making episodes like this, you can support the show at buy me a coffee dot com slash fexingo. That's it. No tiered memberships, no ad reads. Luna: And now back to the solder joint. So what's the most promising approach to bridge this gap? Better sensors or better algorithms? Lucas: Most researchers I've talked to say the bottleneck is sensing, not software. You can have the world's best reinforcement learning model, but if the robot can't feel the temperature of the pad it's trying to solder, it's flying blind. There are some interesting developments in embedded thermocouple arrays inside the soldering iron tip itself — so the sensor is literally at the point of contact. That could give the robot millisecond-scale thermal feedback. Luna: So the iron becomes the sensor. That's elegant. Lucas: Exactly. And a few companies are also experimenting with acoustic sensing — listening to the sound of the solder as it melts and solidifies. The pitch changes based on the wetting angle and cooling rate. It's not mainstream yet, but it's a clever way to get indirect thermal information without adding a camera. Luna: But even with perfect sensing, there's still the dexterity problem. The robot has to feed solder wire at exactly the right rate while moving the iron. That's two coordinated motions on a millimeter scale. Lucas: Right. And most robot arms are designed for positioning, not for applying a controlled feed force. So you either need a custom end effector that integrates the wire feeder, or you teach the robot to use a separate solder dispenser — which adds another degree of freedom and another calibration step. Luna: It's almost like soldering is a microcosm of everything that makes robotic manipulation hard: variable geometry, real-time sensing, multi-step coordination, and the need for near-perfect reliability. Lucas: It really is. And that's why I think the next big leap in electronics assembly won't come from incremental improvements to existing arms. It might require a fundamentally different approach — maybe a soft robotic hand that can hold the iron and sense temperature through its skin, or a learning from demonstration system that watches a human solder a hundred joints and then generalizes the strategy. Luna: But that generalizing part is still the holy grail. We've been hearing about imitation learning for a decade, and it works in demos but rarely in production. Lucas: True. And in the meantime, humans will keep soldering. Which, honestly, isn't the worst outcome — soldering is a skilled trade, and the jobs pay well. But for high-volume, high-reliability applications, the industry would love to reduce human error. The question is whether the solution will come from robotics or from redesigning the components themselves to be more robot-friendly. Luna: Like press-fit connectors instead of solder joints, or laser welding. Lucas: Exactly. Sometimes the easiest way to fix a robot's limitation is to change the task, not the robot. But that's a whole other episode. Luna: Alright, we'll save that for Episode 127. For now, I think we've given listeners a real sense of why that soldering iron is still in human hands. Lucas: And maybe next time you see a circuit board, you'll appreciate the hands that touched it.