Latest / The Robotics Podcast with Fexingo: Autonomous Systems, Industrial Robots, and Hardware / Why Robots Keep Dropping the Screwdriver
Transcript
- Lucas: Let me start with a moment that cost a German automotive supplier eighteen thousand euros. On a transmission assembly line in early 2025, a robot picked up a torque wrench, moved it to the next station, and as it accelerated the arm, the wrench rotated in the gripper by about twelve degrees. The robot placed it crooked, the next operation cross-threaded a bolt, and the entire transmission housing had to be scrapped. Luna: Twelve degrees — that's barely a flick of the wrist. And the robot had force-torque sensors on the wrist, right? Lucas: It had a six-axis force-torque sensor, a high-end model from ATI, the kind that costs about eight thousand euros on its own. The gripper itself was a two-jaw parallel servo-electric with position and current feedback. By the numbers, that cell should have been capable. And yet a human apprentice can pick up that same torque wrench, feel the balance point, and adjust grip pressure instinctively in under a second. The robot couldn't. Luna: So the core problem isn't just 'the robot dropped it.' It's that the robot never truly knows it's losing grip until it's too late. Lucas: Exactly. Think about what 'grip' means for a human hand. We have four type of mechanoreceptors in our skin — Merkel cells, Meissner corpuscles, Pacinian corpuscles, Ruffini endings — each tuned to different frequencies and pressure ranges. When you hold a screwdriver, your brain is integrating signals from all of them in real time. A robot gripper typically has at most one or two strain gauges per jaw, sampling at maybe a hundred hertz, and it's measuring only normal force. Luna: No shear, no slip direction, no texture. So a vibration that's, say, fifty hertz from the arm accelerating — that's right in the middle of what a Pacinian corpuscle would catch — but the robot's sensor just sees noise. Lucas: That's precisely the gap. In the automotive case, the post-mortem showed the arm was running a programmed acceleration profile of about 4 meters per second squared. That created a tangential force on the tool handle. The gripper was applying a normal force of about forty newtons — which should have been sufficient for a two-kilogram tool. But because the center of mass of the torque wrench is offset from the gripper's center, the rotational moment was enough to overcome static friction. Luna: So the robot effectively levered the tool right out of its own grip. And the sensor data logged a force spike that the control logic interpreted as an impact with something — not a slip. Lucas: That's the pernicious thing. The force-torque sensor sees a change in the wrench vector and assumes the tool contacted a fixture. So the robot stops, enters a recovery routine, tries to reorient — and by the time it realizes the tool is now at a wrong angle, the damage is already queued up. The scrap happened not because the robot malfunctioned, but because it didn't have a model of its own grip integrity. Luna: Engineers have been trying to fix this with compliance. Soft grippers, underactuated hands, Fin Ray grippers, even the newer Gecko-inspired adhesives. Do any of those actually solve the tool-handling edge case? Lucas: Partially. A compliant gripper conforms to the tool shape, which increases contact area and therefore friction. In lab tests, a Fin Ray gripper can hold a screwdriver through acceleration profiles up to six meters per second squared — about fifty percent better than a rigid parallel jaw. But the trade-off is positional accuracy. If the gripper deforms, you lose repeatability. For assembly tolerances under fifty microns, that's a non-starter. Luna: So you can either hold it reliably or place it precisely, but not both at the same time. Lucas: That's the engineering frontier right now. There's a group at the German Aerospace Center — DLR — that's been working on what they call 'tactile skins.' These are flexible printed circuit boards with arrays of capacitive tactile cells, each cell measuring normal and shear force at about one kilohertz. They've layered them onto a custom three-finger gripper and shown that the robot can detect incipient slip — the micro-sliding that happens before full loss of grip — within about fifty milliseconds. Luna: Fifty milliseconds — that's fast enough to increase grip force reactively. Did they test it with actual tools? Lucas: They did. In a 2025 paper, they used a standard hex key, a screwdriver, and a ratchet. The tactile skin detected slip events with about ninety-two percent accuracy. But here's the catch: the skin requires a calibration step for each tool. You have to press the tool into the gripper in a specific orientation and run a short training routine. In a production environment where tool changes happen every shift — or where tools get swapped out for maintenance — that calibration overhead becomes significant. Luna: Right, because the whole point of a tool-changer is flexibility. If you have to recalibrate every time you pick up a different screwdriver, you've lost the productivity gain. Lucas: Exactly. And there's another problem: the tactile skin itself is fragile. The DLR prototype lasted about five hundred cycles before delamination started at the edges. For a production robot running thousands of cycles per day, that's not acceptable. Luna: So what about vision-guided regrasp? Some systems now use a wrist-mounted camera to inspect the tool angle after pickup and adjust if it's off. Lucas: That works for detecting a misalignment after the fact. But it doesn't prevent the slip. And if the tool has rotated inside the gripper, the regrasp routine has to put it down and pick it up again — that adds about three seconds per cycle. On a line running a thirty-second cycle time, that's a ten percent hit to throughput. Most plant managers won't accept that. Luna: So the industry is stuck. Rigid grippers slip, compliant grippers lose precision, tactile skins are too delicate, and vision is too slow. Lucas: There is one approach that's showing promise, and it comes from a startup called RoboGrip out of Munich. They've designed a gripper that combines a rigid inner jaw for precision with a outer layer of magnetorheological fluid — a fluid that stiffens in a magnetic field. The idea is that the gripper stays compliant during initial contact, then a quick pulse of current stiffens the fluid to lock the tool in place. They claim they can achieve a coefficient of friction over one point two — compared to about zero point four for a standard rubber pad. Luna: That's a tripling of friction. But magnetorheological fluids require power and control. How fast can they switch? Lucas: The fluid responds in about ten milliseconds. The total cycle is pickup, compliant contact, current pulse, then move. The startup says they've run it through ten thousand cycles on a ratchet without measurable wear. But they haven't published independent validation yet, and they won't say how much the unit costs. Early speculation puts it at around fifteen thousand euros — almost double a standard servo gripper. Luna: So it's expensive and unproven at scale. But if it works, it could close that last ten percent of edge cases that currently cause the scrap events. Lucas: That 'last ten percent' is exactly the problem. In the automotive case, the supplier reviewed a year's worth of production data and found that tool-handling failures accounted for only about two percent of all robot downtime events. But those events caused thirty percent of total scrap cost, because they happened right at the point of assembly. So even a small improvement in grip reliability has outsized financial impact. Luna: It's the asymmetric cost of the edge case. Most of the time the robot handles the tool fine. But when it fails, it fails catastrophically. Lucas: And that's fundamentally a sensing and modeling problem. The robot doesn't have an internal representation of 'the tool is beginning to slip.' It only knows that after the fact, from the torque sensor reading. Until researchers can give robots the equivalent of a slip-detection reflex — something as fast and low-level as the human spinal reflex — we're going to keep seeing these events. Luna: Speaking of things that keep the show running — if today's deep dive into robot grip actually gave you a new way to think about a problem you've seen on your own line, or just satisfied that engineering curiosity, the reason we can keep doing episodes like this without any ad breaks is listener support. It's straightforward: buy me a coffee dot com slash fexingo. That fifty millisecond slip detection fact — that came from a paper we spent time tracking down. If that kind of digging is useful, that's where you can keep it going. Lucas: And it genuinely helps. We don't run ads, we don't take sponsor money, so every episode is just the actual content. The coffee page is the only thing that pays for the research time. Luna: Exactly. So back to the grip problem — one thing I find interesting is whether the automotive supplier changed their process after that eighteen thousand euro scrap event. Lucas: They did. They stopped using the robot for that particular torque wrench pickup. They went back to a manual station for that one step. The robot does everything else. But that one exception — moving the tool from a fixture into the robot's hand — is now a human operation. It's a reminder that even in a highly automated plant, there are still moments where human tactile intelligence is irreplaceable. Luna: And that's not a failure of automation. It's a realistic boundary. The question is: will the next generation of tactile sensors shrink that boundary further? Lucas: That's the million-dollar question — or in this case, the eighteen thousand euro question. I think magnetorheological grippers, combined with embedded shear sensing at the contact surface, could get us there. But we need to see it work in a real production environment, not just a lab bench. Until then, the torque wrench stays with the human. Luna: And maybe that's okay. Automation doesn't have to do everything. It just has to do the things it does well, reliably. Lucas: Exactly. And knowing where the boundary is — that's the actual engineering insight.