Latest / The Robotics Podcast with Fexingo: Autonomous Systems, Industrial Robots, and Hardware / Why Robot Arms Still Cant Pick a Lock
Transcript
- Lucas: There's this video from last December that's been making the rounds in the robotics research community. It's from a group at MIT CSAIL, and it shows a dexterous robotic hand — a Shadow Hand — trying to pick a simple pin-tumbler lock. Luna: Like a padlock you'd buy at a hardware store? Lucas: Exactly. A Master Lock No. 3 — the most common padlock in America. And the robot fails. A lot. The researchers ran a hundred attempts, and the hand succeeded only 23 times. By comparison, a novice human picker — someone who'd practiced for about ten hours total — succeeded 78 times out of a hundred. Luna: Twenty-three percent. That's worse than most people would guess. I think we've gotten used to robots solving Rubik's cubes in under a second. Lucas: Right, and that's the heart of it. Lock picking looks like a fine-motor task — you insert a tension wrench, apply a small torque, then use a pick to lift each pin until it reaches the shear line. But the hard part isn't the positioning; it's the sensing. Luna: The feel. That tiny click when a pin sets, the feedback through the wrench. Lucas: Exactly. In the paper, the researchers note that the Shadow Hand's position control is excellent — it can place the pick tip within a fraction of a millimeter. But torque control at very low forces — we're talking milli newton meters — is still extremely noisy. The robot can't reliably distinguish between the friction of the pick sliding against the pin and the actual 'set' event. Luna: So it's not a strength problem. A robot can easily break the lock. It's a sensory resolution problem. Lucas: Precisely. And this is part of a broader pattern. If you look at the episodes we've done on wet towels, raw eggs, single grains of rice — the common thread is that everyday manipulation requires this blend of force, texture, and slip detection that current tactile sensors just don't deliver at the necessary bandwidth. Luna: Bandwidth is a good way to put it. The human fingertip has something like 2,500 mechanoreceptors per square centimeter. A typical robot fingertip might have a dozen taxels — tactile pixels — and they update at maybe a kilohertz. Human nerves update at several times that. Lucas: And the processing is different. In the lock-picking benchmark, the MIT team used a learning-based approach — they trained a policy in simulation with domain randomization, then transferred it to the real hand. But the simulation couldn't model the subtle acoustic and haptic cues that human pickers rely on. Luna: The click. There's actually a paper from 2023 that showed experienced lock pickers rely heavily on auditory feedback — the sound of the pin hitting the shear line. It's not just feel. Lucas: That's fascinating. So the robot is missing an entire sensory modality. The MIT robot had no microphone. It was purely force-torque. And the policy tried to compensate by using the position of the pick relative to the lock body as a proxy for pin state, but it's not reliable. Luna: Because pins can be set at slightly different heights due to manufacturing tolerances. The lock itself is imprecise. Lucas: Right. And this gets to a deeper point about why robotic manipulation is still so hard. In industrial settings, you can engineer the environment — parts come in precisely oriented trays, with known tolerances. But a lock, a wet towel, a raw egg — these are unstructured, inconsistent, and they require real-time adaptation. Luna: So what's the path forward? More sensors? Better algorithms? Lucas: Both. The MIT group is already working on a second version that integrates a high-bandwidth tactile sensor array from a company called SynTouch — the same sensors used in some surgical robots. They claim a spatial resolution of about one millimeter and a force sensitivity of a few millinewtons. And they're adding a small microphone near the lock. Luna: So basically giving the robot human-like senses. That's the direction of a lot of dexterous manipulation research right now. Lucas: Yeah. And there's commercial pressure too. Companies like Tesla and Amazon are pouring money into general-purpose robot hands for warehouse and home tasks. If you can't pick a lock, you definitely can't fold laundry or handle an egg without breaking it. Luna: It's a benchmark. A simple, cheap, repeatable test that reveals the gap between current capability and human-level dexterity. Lucas: Exactly. One of the paper's authors told me that they chose lock picking specifically because it's a 'gate' skill — if you can do it, you can do a lot of other things. And if you can't, you're stuck in the world of rigid, deterministic tasks. Luna: So where are we now? Two-thousand-twenty-six. Would you bet on a robot picking a lock reliably in the next five years? Lucas: I think within three years, a research lab will demonstrate a hand that can pick a random Master Lock with 90% success in a controlled setting. But generalizing to different lock models, worn locks, locks with dust or corrosion — that's a longer road. Luna: And that generalization is exactly what you need for a robot that cleans your kitchen or unpacks your dishwasher. Each new environment is a new lock, in a way. Lucas: Right. And that's why this paper matters. It's not about lock picking. It's about a reproducible benchmark for force-based manipulation. The MIT team released all their code, the 3D-printed adapters for the lock, and the training environment. So now other labs can iterate. Luna: That's the kind of open science that drives progress. And on that note — if this episode gave you a better sense of why robot hands still struggle with everyday objects, and if it felt worth a coffee to you, there's a link. It's buy me a coffee dot com slash fexingo. No pressure, just a small way to keep the show ad-free and focused on stories like this. Lucas: Yeah, that's it. The smallest possible ask. And we appreciate it. So — back to the lock. One thing that surprised me in the paper was how much the robot struggled with the torque direction. Human pickers naturally vary the torque — they apply a slight counterclockwise pressure, then release a tiny bit to let a pin drop. The robot's policy was trained to apply a fixed torque, and it would get stuck. Luna: Because a fixed torque doesn't account for the varying spring resistance of different pins. Some pins are stiffer than others, even in the same lock. Lucas: Exactly. And the robot couldn't learn to modulate because it couldn't feel the subtle change in resistance. So the next step is to combine the high-resolution tactile sensors with a control policy that can handle the partial observability — basically, the robot has to infer the internal state of the lock from limited feedback. Luna: That sounds like a classic partially observable Markov decision process. POMDP. Lucas: Exactly. The researchers framed it that way in a follow-up preprint. And they showed that with a POMDP solver, even with the same noisy sensors, they could improve success to 41 percent. Still not great, but nearly double. Luna: So better inference compensates for crappy sensors. That's promising. Lucas: It is. And it suggests that we might see a hybrid approach win — better sensors plus better algorithms. The question is whether the cost comes down enough for commercial robots. A SynTouch sensor array for one fingertip costs about two thousand dollars. That's more than many robot hands themselves. Luna: But prices drop with scale. If Amazon decides it needs tactile sensors for its warehouse robots, the unit cost could fall by an order of magnitude within a couple of years. Lucas: That's the bet. And we're already seeing it with force-torque sensors at the wrist — they used to be a thousand dollars, now you can get a decent one for under two hundred. So there's a clear trajectory. Luna: So the lock-picking robot is a canary in the coal mine. It tells us how far we have to go in tactile sensing and adaptive control. Lucas: Exactly. And it's a concrete, replicable test. Next time you see a robot hand doing something impressive on YouTube, ask yourself: can it pick a Master Lock No. 3? If not, we're still a few steps away from the general-purpose home robot. Luna: I'm going to start using that as my personal benchmark. Lucas: Good. And if you want to try it yourself — the MIT code is public. You could build a robot hand and test it. Though I'd warn you, it's harder than it looks. Luna: Or you could just buy a lock and pick it manually. That's probably easier. Lucas: Probably. And a lot cheaper. But not as cool. Luna: True. Alright, that's the state of robotic lock picking in twenty twenty-six. Next episode, we'll look at a different manipulation challenge.