Latest / The Robotics Podcast with Fexingo: Autonomous Systems, Industrial Robots, and Hardware / Why Robot Arms Still Can't Trust Their Own Force Sensors
Transcript
- Lucas: So here is a number that stopped me this week. A major EV battery plant reported a 12 percent scrap rate on one of its module assembly lines, and the root cause was not bad welds or contaminated cells — it was the robot's own force sensor giving false readings. Luna: Wait, the robot was scrapping good parts because it thought something was wrong? Lucas: Exactly. The arm was pressing a busbar into a terminal, and the force-torque sensor at the wrist would report a spike that looked like a misalignment — so the robot would abort the cycle and flag the part as defective. But the part was fine. The sensor was just drifting. Luna: And a quick honest thing — we're able to talk about stuff like this without any sponsor breaks because a small group of listeners chips in through buy me a coffee dot com slash fexingo. That's literally what keeps us ad-free and allows us to spend time on these deep dives. Lucas: Yeah, it's a big help. So back to that sensor drift — the battery plant had to pull a technician offline every two weeks to recalibrate the sensor manually. That's lost production, lost labor, and 12 percent of their output in the dumpster. Luna: And force sensors are not cheap. A decent six-axis force-torque sensor from ATI Industrial Automation can run five to ten thousand dollars. So the plant wasn't using junk. It was high-end hardware. Lucas: Right. And the problem is fundamental. Most force sensors rely on strain gauges — tiny resistors that change resistance when the metal they're bonded to deforms. But those gauges are sensitive to temperature, to humidity, to the adhesive aging. Over a shift, the zero point can wander by a few tenths of a Newton. Luna: A few tenths of a Newton — that's the weight of an egg. So if your tolerance window is plus or minus half a Newton, a drift of 0.3 Newtons could push you right to the edge. Lucas: And that's exactly what happened. The sensor thought it felt a 1.7 Newton force when the real contact force was 1.4. The robot's control loop saw the number, decided the part was jammed, and stopped. Over a thousand cycles, that false abort happened about 12 percent of the time. Luna: So why can't the software just filter out that drift? If the sensor always reads a little high, can't the controller subtract the offset? Lucas: It can — and most systems do run an auto-zero at startup. The robot moves to a known position with no contact, measures the baseline, and subtracts it. But the problem is that drift is not constant. It changes with temperature, with the arm's orientation, with how long the robot has been running. An auto-zero at 7 AM might be off by lunch. Luna: And that's why some factories have resorted to what's called 'touch and learn' calibration. The robot gently touches a known surface — like a ground steel block — and records the expected force profile. Then it uses that as a reference for the rest of the shift. Lucas: But even that can fail if the block gets a nick, or if the robot's mounting compliance shifts. The whole system is only as good as the mechanical reference. And in a high-volume battery line, you can't stop every thirty minutes to retouch a block. Luna: So what are the alternatives? I know some newer sensors from Robotiq use capacitive sensing instead of strain gauges. Does that help? Lucas: Capacitive sensors measure changes in capacitance between two plates as the force deforms a dielectric layer. They tend to have better temperature stability — drift is lower. But they're more expensive, and they're still susceptible to long-term creep. The dielectric can take a set over thousands of cycles. Luna: So every force sensor technology has a failure mode. What about optical sensors? I've seen papers on using fiber Bragg gratings — where the wavelength of reflected light shifts with strain. Lucas: Optical sensors are incredibly accurate — microstrain resolution — and they're immune to electromagnetic interference. But they require a laser interrogator that costs tens of thousands of dollars. For a single robot, that's hard to justify. And the fiber itself is fragile at the wrist — it can crack when the robot makes sudden moves. Luna: So we're stuck with strain gauges for most applications, and we're stuck with drift. Has anyone tried using AI to predict the drift and compensate in software? Lucas: Yeah, a few groups have. You can train a neural network on historical temperature, humidity, and cycle data to estimate the sensor offset in real time. It works in the lab — I saw a paper from MIT CSAIL where they cut drift error by 80 percent. But in production, the model can fail if the environment changes in a way it wasn't trained on. Luna: Right, like a new HVAC system changes the airflow pattern, or a nearby machine starts throwing off heat. The model suddenly sees inputs outside its training distribution and makes bad predictions. Lucas: Exactly. So the conservative approach is still manual calibration cycles. Some factories have automated that — they program the robot to touch a known surface between every, say, 50 cycles. That adds maybe 2 seconds per cycle, but it keeps the scrap rate below 1 percent. Luna: Two seconds per cycle — that's still a huge hit on throughput. If your cycle time is 30 seconds, you're losing nearly 7 percent of your capacity just to calibration. Lucas: And that's the trade-off. You either accept the scrap rate, or you eat the cycle time. And in high-margin industries like medical device assembly, they accept the cycle time because a false positive is better than a false negative. But in automotive battery lines, margins are thin, so they push the tolerances and accept the scrap. Luna: It reminds me of the old dilemma in statistical process control — you can't have both high precision and high speed without paying for it. And here, the sensor itself is the weak link. Lucas: The hardware is getting better. Some new sensors use multiple redundant strain gauges — like six instead of four — and take a median reading. That filters out a single gauge's drift. But it adds cost and complexity in the signal conditioning. Luna: And then the robot controller has to handle six channels instead of four, which means more wiring, more connectors, more potential failure points. Lucas: Right. It's not a simple fix. I think the real breakthrough might come from a different approach entirely — force estimation without a dedicated sensor. Some research groups are using the motor currents from the robot's joints to infer the external force. No sensor at the wrist, no drift. Luna: Motor current sensing is already used for collision detection in collaborative robots. But the resolution is lower — you can't feel a 0.1 Newton force through the gearbox friction and inertia. Lucas: That's the challenge. But with high-resolution encoders and better friction models, it might be possible. A few years ago, a team at ETH Zurich demonstrated a controller that could detect forces down to 0.5 Newtons using only joint torque sensors — which are essentially strain gauges on the gearbox output, but they're more stable because they're not at the wrist. Luna: So the wrist sensor is the biggest problem because it's the most exposed. It sees temperature changes from the end effector — if you're picking up a hot part from a curing oven, the sensor heats up and drifts. Lucas: Exactly. And the solution might be to move the sensing away from the wrist entirely. Use the base torque sensors for force control, and only use the wrist sensor for high-frequency vibration sensing — like detecting chatter in a grinding operation. Luna: So we're back to the idea of putting the sensor where the physics is friendlier. That's a design choice, not a technology problem. Lucas: And it's a choice that most robot manufacturers haven't made because it requires rethinking the whole control architecture. A lot of legacy systems assume a wrist sensor is the only way to do force control. Changing that is a multi-year engineering effort. Luna: So for now, the factory floor is stuck with drift, scrap, and the occasional technician with a torque wrench and a laptop running calibration scripts. Lucas: Pretty much. And the irony is that a human can feel a 0.1 Newton difference with their fingertips. A robot with a $10,000 sensor cannot. That's the gap we're still trying to close. Luna: And on that note, if today's conversation gave you something to think about, the simplest way to keep this show ad-free is through buy me a coffee dot com slash fexingo. No pressure, just wanted to mention it. Lucas: Yeah. So next time you see a robot arm on a factory floor, just remember — it might be lying to itself about how hard it's pushing. And that's where the real engineering work begins.