Latest / The Robotics Podcast with Fexingo: Autonomous Systems, Industrial Robots, and Hardware / Why Robot Grip Strength Still Confuses a Doorknob
Transcript
- Lucas: So we have spent the last twenty episodes on robot hands, and I think we have covered everything from picking screws to gripping raw chicken. But there is one simple object that still makes them look completely helpless, and it is sitting on every door in the building. Luna: Let me guess: the doorknob. Lucas: The doorknob. And not because the robot cannot reach it. The problem is grip strength. Or more precisely, the complete lack of subtlety in how much force to apply. Luna: So we are not talking about a robot being too weak. We are talking about it being too strong. Lucas: Exactly. You can teach a robot to exert fifty newtons of force on a cylindrical object, which is plenty to hold a soda can. But a doorknob requires a very different approach. You need to apply enough tangential force to overcome the latch, but not so much radial force that you crush the handle or strip the mechanism. Luna: And robots are notoriously bad at that because they think in binary: grip or release. Lucas: Right. Most industrial grippers are designed for pick and place: grip a part, move it, release. They are either fully open or fully closed. There is no in-between. So when a robot tries to turn a doorknob, it either squeezes too hard and jams the mechanism, or it does not squeeze hard enough and the hand slips off. Luna: I remember reading about a home robot competition where a robot literally ripped a door off its hinges because it tried to turn the knob like it was opening a jar. Lucas: That is the classic failure mode. The robot applies, say, a hundred newtons of force radially, which is way more than needed, and the torque that generates is enough to shear the screws. The door comes off. It is a great demo of why raw force is not the answer. Luna: So what is the answer? Is it better sensors, better algorithms, or better mechanical design? Lucas: All three, but let me give you a concrete example. Researchers at a well-known university built a robot hand that uses force sensors in each fingertip to measure both normal and shear forces. They trained it on a dataset of human demonstrations of turning doorknobs, and then they let it loose on a variety of knob types. The key was that the controller did not just target a specific force. It targeted a force profile over time. Luna: So it is not about holding a constant grip, but modulating it in real time. Lucas: Exactly. As the knob turns, the resistance changes. The robot has to increase or decrease the grip force based on the feedback from the sensors. And it has to do that in milliseconds. That is where the challenge lies. Luna: I have also heard about using torque sensors in the wrist to measure the reaction torque. That seems like a simpler approach. Lucas: That works for a fixed axis, but a doorknob is not always coaxial with the robot's wrist. You get coupling effects. The wrist torque sensor sees a combination of the knob's resistance and the misalignment forces. It is messy. Luna: So fingertip force sensing is more direct. Lucas: It gives you the contact point, which is crucial. And modern tactile sensors can measure not just force magnitude but also the distribution across the contact patch. That is how you detect micro-slips before the hand loses grip entirely. Luna: Micro-slips. That is the term. I have seen papers on that. The robot can actually feel the onset of a slip and increase grip slightly to prevent it. That is what humans do unconsciously. Lucas: And that is the key insight. Humans are incredibly good at maintaining grip with minimal effort. We use just enough force to prevent slip, and we adjust on the fly. Robots are only starting to mimic that with high-frequency tactile sensing and fast control loops. Luna: So how far are we from a robot that can reliably open a door? I am not talking about a lab demo, but something you would trust in your home. Lucas: Realistically, we are a few years away. The hardware is getting there, but the software is still the bottleneck. You need to handle different knob geometries, different friction coefficients, and unexpected obstructions like a painted-shut door. It is a perception and control problem, not just a hardware problem. Luna: And that is frustrating because opening a door is such a mundane task for us. We do not even think about it. Lucas: That is exactly what makes it so hard. The things we do without thinking are the hardest to teach machines. And I think that is a deep lesson for anyone building robots: the last mile is not about strength, it is about restraint. Luna: Restraint. I like that. It is almost the opposite of what you would expect from a machine designed to move heavy things in a factory. Lucas: And that is why the next generation of robots for homes and hospitals will not be built like industrial arms. They will be built with compliance, with soft materials, with sensors everywhere. The goal is not to apply maximum force. It is to apply the right force at the right time. Luna: And that ties back to what we have been saying about robot hands learning to feel a screwdriver or judge weight. They are all part of the same story: giving robots the sense of touch. Lucas: Absolutely. And if today's conversation gave you something useful, that is exactly the kind of thing we want to keep doing. This show stays ad-free because of listener support, and if that matters to you, you can help at buy me a coffee dot com slash fexingo. Luna: It is a small gesture, but it keeps the lights on and the robots talking. We appreciate it. Lucas: So back to the doorknob problem. What are the actual numbers? A typical interior door latch requires about ten newtons of force to retract. But the radial grip force needed to turn the knob without slipping depends on the friction coefficient between the robot's fingertip and the knob material. For a smooth metal knob, you might need fifty newtons of normal force to generate ten newtons of tangential force. That is a five-to-one ratio. Luna: And if the robot only knows how to grip at a hundred newtons, it is already using twice the necessary force, which can cause damage or make the knob too hard to turn. Lucas: Exactly. And to make it worse, the required force changes as the knob turns because the mechanical advantage of your hand changes. So you need real-time adaptation. That is what the best research systems are doing now. Luna: I have also seen some clever mechanical solutions, like passive compliance in the fingers. If the finger can flex slightly, it can absorb some of the misalignment without needing to compute it. Lucas: Yes, passive compliance is a big deal. Some groups are using soft materials or spring-loaded joints that naturally yield when the force gets too high. That reduces the need for precise control and makes the system more robust to uncertainty. Luna: So we have two approaches: one is to sense everything and control precisely, the other is to build in mechanical intelligence that does some of the work for free. Lucas: And the best systems combine both. You have a compliant hand that can conform to the object, plus high-resolution sensors that tell the controller what is happening at the contact. It is a nice synergy. Luna: What about learning? Are they training neural networks to do this directly? Lucas: There is a lot of reinforcement learning work where the robot learns a policy by trial and error in simulation, then transfers to the real world. The challenge is the sim to real gap. You have to model the contact dynamics accurately, which is notoriously difficult. But with better tactile simulators, we are seeing progress. Luna: So maybe in a few years, we will have a robot that can open any door without ripping it off the hinges. That is the dream, right? Lucas: That is the dream. And once you can open a door, you can navigate a building, you can perform tasks in different rooms, you can truly be a home assistant or a hospital aide. It is a fundamental skill. Luna: It is funny to think that the humble doorknob is one of the last frontiers for robotics. Lucas: It is. But it is a great reminder that the most everyday objects are often the hardest to master. And that is what makes this field so interesting. Luna: I cannot wait to see the day when a robot opens a door for me. It will be a small step for a robot, but a giant leap for robot kind. Lucas: Ha, exactly. And when that happens, we will dedicate an episode to it. Until then, keep your doorknobs safe from overly enthusiastic robots.