Latest / The Robotics Podcast with Fexingo: Autonomous Systems, Industrial Robots, and Hardware / Why Robot Arms Cant Operate a Doorknob
Transcript
- Lucas: You know, I've been thinking about the last few episodes — wet towels, raw eggs, single grains of rice. And it hit me: there's one task that's so universal, so mundane, that every human does it dozens of times a day without thinking, and yet it's a genuine frontier in robotics. Luna: Let me guess — turning a doorknob? Lucas: Exactly. Opening a door. And I don't mean the sensor-sliding automatic doors at the grocery store. I mean a standard interior door with a lever or a round knob. It sounds trivial, but it's actually a perfect storm of challenges for a robot arm. Luna: I've seen the videos. A robot hovers near the handle, tries to grip it, and either misses or pulls too hard and yanks the whole door off the hinges. Lucas: Right. And those failures reveal something fundamental. The problem isn't just about gripping — it's about what roboticists call 'compliant motion.' When you or I grab a doorknob, we instinctively apply just enough force to turn it while simultaneously feeling the resistance of the latch. We're not moving in a pre-planned path; we're adjusting in real time based on tiny tactile cues. Luna: So the robot needs to both sense force and control its own stiffness? Lucas: Yes. And that's where force-torque sensors come in. These are sensors that measure the forces and torques applied to the robot's wrist. For a door-opening task, you need to be able to feel the interaction forces — the resistance of the latch, the weight of the door, the friction in the hinge. Then you need to control the robot's impedance — basically how stiff or compliant it is — to respond appropriately. Luna: But those sensors have been around for decades. Why is it still so hard? Lucas: The sensors are good, but the control loops have to run at very high frequencies — like a kilohertz or more — to react to sudden changes. And the geometry of the door itself is variable. The handle height, the swing direction, the friction of the latch, even the weight of the door all change from door to door. So you can't just program one trajectory and call it done. Luna: So it's a perception problem too. The robot has to figure out what kind of door it's facing. Lucas: Exactly. And this is where a lot of the recent progress is coming from. In the DARPA Subterranean Challenge, which wrapped up last year, teams had to navigate robots through underground tunnels and buildings. One of the critical bottlenecks was door opening. Most teams used a combination of vision — like a depth camera to estimate the handle pose — and force feedback to execute the grip and turn. Luna: I remember seeing the winning team from Carnegie Mellon — they used a robotic arm with a specialized gripper that could also push the door once it was unlatched. Lucas: Yeah. That's the other piece — the actual manipulation after the latch is released. You have to either push or pull the door open, and the robot needs to switch from a gripping motion to a pushing motion seamlessly. If the gripper is too rigid, it might lose grip. If it's too compliant, it won't have enough torque to turn the handle. Luna: It sounds like a control algorithm problem as much as a hardware one. Lucas: Definitely. And there's been some promising work in learning-based approaches. For instance, researchers at UC Berkeley in 2023 trained a neural network to predict the required forces for turning a variety of door handles, using simulated data and then transferring to a real robot arm. Their success rate on novel handles was something like 85 percent — not perfect, but a huge improvement over earlier efforts. Luna: But that still leaves 15 percent of doors that stump the robot. What kinds of doors fail? Lucas: The tricky ones are usually doors with very stiff latches, or handles that require a combination of push and turn — like some lever handles that need to be pressed down while simultaneously pulled. Also, doors with high friction hinges or those that are warped from humidity. And round knobs are notoriously harder than levers because they require a firm rotational grip without any visual cues for alignment. Luna: So if I have a sticky door in my house, a robot would probably struggle. Lucas: Almost certainly. And that's a big barrier for home assistance robots. The whole promise of a robot that can fetch items or help someone with limited mobility runs into this first hurdle: it has to be able to move between rooms. A robot that can't open doors is confined to a single space. Luna: Are there any commercial robots that actually solve this reliably? I know some of the humanoid robots from Boston Dynamics and others have demonstrated door opening. Lucas: They have, but usually in controlled demonstrations. Boston Dynamics' Atlas robot has opened doors in research settings, but it uses a custom end-effector and often relies on pre-mapped environments. For a production-ready system that can handle any door, we're probably still a few years out. But the progress is real. Luna: It's interesting that something so basic for us is such a benchmark for dexterity. Lucas: It really is. And it's a good reminder that a lot of what we think of as simple motor skills are actually incredibly complex sensorimotor loops. For a robot, even the act of aligning the gripper with the handle requires solving a 3D pose estimation problem, sometimes with reflective or shiny surfaces. Luna: And then you have the door itself — which is a moving, articulated object with unknown friction and mass. Lucas: Right. So it's a microcosm of the broader challenge in robotics: bridging the gap between precise, repeatable tasks in controlled environments, and the messy, variable world humans operate in. If we can solve door opening robustly, a lot of other manipulation tasks become easier. Luna: It's funny — we've spent years talking about why robots can't handle a single piece of spaghetti, but a door is something they'd actually need to handle in the real world. Lucas: Exactly. And there's a real economic incentive. Think about elderly care facilities, hospitals, or even warehouses with manual doors. If a mobile robot can autonomously navigate through a building, it can do deliveries, fetch supplies, and perform inspections. The door is the linchpin. Luna: Speaking of which, I remember a startup called 'DoorBot' — no, that's not the name, sorry. Actually there's a company called 'Levitate' — they were working on a retrofittable door opener for existing buildings, but that's more of an accessory than a robot arm. Lucas: That's a different approach entirely — modifying the environment instead of the robot. And in many industrial settings, that's exactly what happens: automatic doors or magnetic switches. But for truly autonomous robots that need to operate in any building, they have to handle standard doors. Luna: So what's the state of the art right now, in mid-2026? Lucas: We're seeing a convergence of better sensors, more robust control, and learning from demonstration. A few research groups have shown robots that can open doors after just a few trials, using reinforcement learning to adapt to the specific door's dynamics. But robust generalization to unseen doors — across different handle types, latch forces, and door weights — remains an open problem. Luna: It sounds like a great thesis topic for a PhD student. Lucas: Absolutely. And there's even a benchmark called the 'Door Opening Challenge' that some robotics conferences have. Teams compete to see who can open a set of unknown doors the fastest. It's surprisingly competitive. Luna: I'd watch that on YouTube. Honestly, if today was worth a coffee to you, that's the link — buy me a coffee dot com slash fexingo. Lucas: Yeah, listener support really helps us keep digging into these niche but fascinating corners of robotics. And if you want to see those door-opening fails, they're out there — just search 'robot fails at door'. Luna: Anyway, back to the tech. So where do you see this going in the next five years? Lucas: I think we'll see the first commercial service robots that can reliably handle standard door types — maybe not all doors, but a significant majority. As force sensors get cheaper and control algorithms get more data-driven, the cost will come down. And once you can open a door, the range of applications for mobile manipulation explodes. Luna: It's one of those milestones that, once achieved, will seem obvious in hindsight. Like when phones got touch screens. Lucas: Exactly. And then we'll wonder why it took so long to teach a robot to do something a two-year-old can do.