Latest / The Robotics Podcast with Fexingo: Autonomous Systems, Industrial Robots, and Hardware / Why Robot Arms Still Cant Pick a Lock
Transcript
- Lucas: If today's tech conversation gave you something usable — a new way of thinking about a problem you actually bump into — there's a quiet reason for that. A handful of listeners chip in monthly to keep this show ad-free, and that's what makes episodes like this one possible. Luna: Yeah, it's a small group, but it adds up. And it means we never have to tailor the content for ad-friendly angles. Lucas: Exactly. So if you've gotten something out of the show — maybe a clearer picture of why your robot vacuum still bumps into the same chair — and you'd like to be part of that, it's buy me a coffee dot com slash fexingo. No pressure, no perks, just knowing you're helping keep this independent. Luna: And we're back to the hardware. So Lucas, you were saying that picking a lock is absurdly hard for a robot? Lucas: Right. Let's start with something most of us have done — picking a simple pin and tumbler lock. You slide in a tension wrench, apply a tiny torque, then feel for the pins binding. It's all about force feedback under a gram of pressure. And that's exactly where robot arms fail. Luna: We've done a whole series on why robot arms can't handle stuff like wet soap or a paperclip. Is lock picking just another example of that same problem? Lucas: It's the same core issue — tactile sensing — but lock picking is a particularly pure test. Because it's not about grasping an object's shape; it's about detecting a barely perceptible change in resistance. A human can feel a pin 'click' when it reaches the shear line. That click involves a force change on the order of millinewtons. Luna: Millinewtons. So even high-end industrial force-torque sensors are way too coarse? Lucas: Way too coarse. A typical six-axis force-torque sensor on a robot like the Fanuc LR Mate 200iD has a resolution around 0.1 newtons. That's a hundred millinewtons. You need something ten to a hundred times more sensitive. And you need it in a tiny package that fits inside a gripper. Luna: Plus you need to apply that torque while simultaneously feeling the feedback. That's a control loop challenge too, right? Lucas: Exactly. In a human hand, the muscles that apply the tension are the same ones that sense resistance — it's all integrated. In a robot, the torque is applied by motors, and the sensing is done by separate sensors. There's a time lag and a signal to noise problem. And the feedback loop has to run at kilohertz speeds to catch that millisecond 'click'. Luna: So even with a sensitive sensor, the robot might miss the shear line entirely because it's not reacting fast enough. Lucas: That's exactly what happened at the 2025 LockPicking Robot Challenge. Researchers from the University of Bristol entered a robot arm with a custom three-fingered gripper, each finger equipped with a BioTac tactile sensor — that's a biomimetic sensor that mimics human skin. They could detect forces as low as 0.01 newtons. Luna: That's ten millinewtons — so sensitive enough in theory. Did it work? Lucas: It picked a four-pin lock in about three minutes. But it failed on a five-pin lock because the accumulated tolerance stack — each pin's shear line is slightly different — confused the control algorithm. The robot couldn't distinguish between a pin that was set and one that was over-set. Luna: Over-setting — that's when you push a pin too far and it drops into the next chamber, right? Lucas: Right. And a human would feel that — the tension wrench suddenly gets loose — and they'd back off. The robot's algorithm didn't have a model for that failure mode. It just kept pushing. Luna: So it's not just sensing — it's also the control software being able to interpret ambiguous tactile signals. Lucas: That's the second big problem. Humans learn through thousands of hours of practice. A lock picker builds a mental model of how a pin feels when it's binding versus when it's set. A robot has to learn that from data. And there's not a lot of high-quality labeled data for lock-picking forces. Luna: You could simulate it, but the physics of pin tumbler locks is pretty complex — friction, wear, different keyways. Lucas: Right. And the third problem is the tool itself. A robot gripper holding a tension wrench isn't very dexterous. The wrench has to be angled exactly, and the torque applied must be precisely perpendicular to the keyway. A slight misalignment and the wrench slips or breaks. Luna: So even if you had perfect sensing and perfect control, the mechanical interface — the hand — is still clumsy. Lucas: Exactly. That's why some researchers are moving away from traditional industrial grippers and toward soft robotic hands with multiple degrees of freedom. A soft hand can conform to the shape of the wrench and apply torque more naturally. Luna: That's reminiscent of the work at Carnegie Mellon on the 'morphing hand' that can do a power grip and a precision grip. Lucas: Yes. But even those hands are still far from human-level dexterity. The LockPicking Robot Challenge really highlighted that we're maybe five to ten years away from a robot that can pick a standard pin-tumbler lock reliably — and that's assuming big advances in tactile sensing and learning algorithms. Luna: But why would we want robots that can pick locks? It seems like a niche skill. Lucas: It's a benchmark — like the Rubik's cube for manipulation. Lock picking tests force control, precision, and real-time adaptation. But there are also practical applications: security testing, emergency response, and even automated assembly of delicate components that require 'feel'. Luna: And if a robot can pick a lock, it can probably also insert a plug into a socket, or align a tiny screw — tasks that currently stump most assembly robots. Lucas: Exactly. So the lock-picking problem is really a proxy for a whole class of manipulation tasks that require gentle, adaptive force control. And we're not there yet. But every failed attempt teaches us something about what's missing. Luna: So what's the most promising approach you've seen? Lucas: I'd point to the work at the University of Tokyo on 'tactile servo' control. They integrate a high-resolution tactile sensor directly into the gripper's fingertips, and run a control loop at four kilohertz. They can detect slips of less than ten micrometers. That's the kind of speed and sensitivity needed to feel a pin click. Luna: And they've demonstrated picking a lock? Lucas: Not yet publicly. But they've succeeded at tasks like inserting a key into a lock and turning it — which is actually a different skill. Insertion requires alignment; picking requires feeling internal states. So it's a step. Luna: I wonder if we'll ever see a robot that can pick a lock faster than a human. That seems like a long shot. Lucas: Maybe not faster, but more consistently. A human lock picker can have good days and bad days. A robot, once it learns the skill, could pick any lock of the same type with perfect repeatability. That's valuable for security audits — imagine a robot that can test every lock in a building overnight. Luna: That would keep locksmiths in business, I suppose. But it also raises some ethical questions. Lucas: It does. Any technology that can bypass security can be misused. But the same could be said for lock picking tools themselves. The research community is aware of that, and there's a code of conduct around publishing lock-picking robot designs — you don't release the exact control parameters that would make it easy to replicate. Luna: So it's like the 'weaponization' debate in AI — responsible disclosure. Lucas: Exactly. And the LockPicking Robot Challenge itself is a controlled environment; the organizers provide the locks and the tools. It's not like anyone's building a black-bag robot. Luna: Alright, so to wrap up: robot arms can't pick a lock because they lack the tactile sensitivity, the control speed, and the dexterous tool manipulation. But researchers are getting closer, and the benchmark is driving real advances in force sensing. Lucas: That's the story. And as those advances trickle down, we'll see better assembly robots, better surgical robots, maybe even robots that can finally fold that fitted sheet. But for now, your front door is safe from the robot apocalypse. Luna: Good to know. I'll keep my key in my pocket. Lucas: And I'll keep practicing with my practice lock. Maybe I'll have a robot student someday.