Latest / The Robotics Podcast with Fexingo: Autonomous Systems, Industrial Robots, and Hardware / Why Robot Arms Still Can't Open a Door
Transcript
- Lucas: So we've spent over twenty episodes now cataloging the things robot arms cannot do. Peel potatoes, open jars, tie knots, thread needles — the list is both humbling and fascinating. Luna: And I think listeners might be wondering — have we run out of everyday objects that stump a robot? Turns out, no. Lucas: Not even close. Today's object is something you interact with dozens of times a day without thinking. A door. Luna: A door? Seriously? I've seen robot vacuums that can open doors. Lucas: You've seen robot vacuums that nudge a door that's already slightly ajar. I'm talking about a closed door — a push-bar door, a lever handle, a round knob, a sliding door — the full spectrum. And the research shows that even the most advanced industrial arms, the kind that can assemble a transmission with sub-millimeter precision, fail at this task roughly half the time. Luna: Half the time? That's worse than my four-year-old nephew. Lucas: And your nephew doesn't have a six-axis arm with force-torque sensing and a vision system. So why is a door so hard? The short answer: a door is what roboticists call a 'contact before constraint' problem. Luna: Let's unpack that phrase because I think it gets at the core of why so many manipulation tasks are hard. Lucas: Right. Most industrial robot tasks are 'constraint before contact'. The robot knows exactly where the part is, it moves to that exact coordinate, and then it makes contact. The constraint — the fixture, the jig, the known geometry — comes first. A door is the opposite. The robot has to make contact with the handle first, then figure out the constraint. The handle could be at any height, any angle, any force profile. And the door itself — the hinge, the weight, the friction — all of that is unknown until the robot starts pushing. Luna: So it's not just about grabbing the handle. It's about understanding the entire dynamic system in real time. Lucas: Exactly. And there's a brilliant 2025 paper from MIT's CSAIL — led by a researcher named Naomi Hersch — that tried to formalize this. They built a door dataset with over 200 different door types in real buildings: offices, labs, warehouses, hospitals. Then they trained a reinforcement-learning policy on a KUKA arm with tactile sensors on the gripper. Luna: And the results? Lucas: On doors the arm had seen during training — what they called 'familiar door types' — it succeeded about 85 percent of the time. On doors it had never encountered, success dropped to 52 percent. Luna: So basically, it's memorizing specific door mechanics, not learning a general 'open door' skill. Lucas: That's the current ceiling. The arm can learn the force profile of a specific lever handle in a specific building, but transfer that knowledge to a push-bar in a different building? Almost zero transfer. And this is where we get to the deeper issue — semantic interaction. Luna: Semantic interaction meaning the robot doesn't just see a shape; it understands that this object is meant to be pushed down, or pulled, or turned, based on context. Lucas: Yes. Humans do this effortlessly. You walk up to a door and you don't even look at the handle; you just know from the presence of a vertical bar that it's a push-bar. You know from a round knob that you need to twist. A robot's vision system can classify the handle type — it can say 'this is a lever' with 99 percent accuracy — but that classification doesn't translate into the correct motor plan. The motor plan has to be discovered through physical interaction, which takes time and risks failure. Luna: And failure with a door isn't just dropping an object. A robot that pushes a push-bar too hard can break the mechanism. A robot that twists a round knob too fast can strip the internal latch. There's actual damage involved. Lucas: That's the real-world cost. There's a well-known incident from 2023 where a delivery robot from a major autonomous vehicle company — I won't name them — got stuck at a hospital lobby door for forty-five minutes. It kept trying to pull a push-bar. It eventually had to be rescued by a human. The video went viral inside the robotics community as a cautionary tale. Luna: And this isn't just a lab curiosity. For warehouses, hospitals, hotels — anywhere you want a robot to navigate between rooms autonomously — doors are the final barrier. You can have the best navigation system in the world, but if you can't open the door, you're stuck. Lucas: Speaking of which, I want to take a quick moment here — because this conversation about real-world robotic failures is exactly the kind of deep, technical dive that keeps this show ad-free. If you're getting value from episodes like this, and you want to help us keep producing them without sponsors, listener support is what makes it possible. For those interested, you can find us at buy me a coffee dot com slash fexingo. It's a very simple way to chip in, and it genuinely helps us stay independent. Luna: Yeah, and it's something we're grateful for. Anyway, back to doors — the Hersch paper also highlighted something else: the door handle itself is a surprisingly complex mechanical interface. Lucas: It is. They measured the force required to operate over 150 different handles. A lever handle in a modern office building might require only 5 Newtons of force. A push-bar on a fire door can require 30 Newtons. And that's just the static force — the force to start the motion. The dynamic force changes as the latch retracts and the door swings open. The robot has to continuously adjust its grip and its force output in real time. Luna: And the geometry varies too. Some levers are short, some are long. Some push-bars are horizontal, some are angled downward. Some knobs are smooth, some are textured. It's an infinite combinatorial space. Lucas: That's exactly why a purely geometric approach fails. You can't pre-program every possible door. The only path forward seems to be learning-based, but with much better generalization. Some groups are exploring what's called 'meta-learning' — training a policy that can adapt to a new door in just a few seconds of interaction. Luna: So the robot tries the door, fails, learns from the failure, and tries again — all within a couple of attempts. Lucas: Exactly. The Hersch lab actually built a system that does this. On an unseen door, the robot first performs a 'probing' motion — it gently pushes the handle in several directions to measure compliance. Then it uses that data to update its internal model of the door and generates a new motor plan. That improved success on unseen doors from 52 percent to 68 percent. Luna: Still not great, but it's progress. And it's interesting that the probing motion itself is something humans do without thinking. You jiggle the handle slightly before committing to the full motion. Lucas: Right. That jiggle is information gathering. The robot has to learn to do the same thing. And that requires a tight loop between tactile sensing, motor control, and decision-making. It's not just an engineering problem — it's a control theory problem, a machine learning problem, and a mechanical design problem all at once. Luna: I want to ask about the hardware side. Are there grippers specifically designed for door handles? Lucas: There are, and they're surprisingly varied. Some labs use parallel-jaw grippers with soft rubber pads. Others use three-fingered hands. One interesting approach from a startup called Flexiv uses a 'fin-ray' gripper — inspired by fish fins — that passively conforms to the shape of the handle. The idea is that if the gripper can adapt its shape without active control, the planning problem becomes simpler. Luna: So the gripper itself does some of the work that the software would otherwise have to do. Lucas: Exactly. It's a classic embodied intelligence approach: offload complexity to the hardware. But even the best gripper can't solve the fundamental problem of unknown dynamics. If the door is unusually heavy or the hinge is sticky, the gripper doesn't know that until it's already committed. Luna: And I imagine doors also present a safety issue. If a robot misjudges and applies too much force, it could injure a person on the other side. Lucas: Absolutely. That's the other reason this problem has been slow to solve. In industrial settings, robots are often behind cages. But a door-opening robot is interacting with human spaces. It has to be compliant enough not to harm anyone, but firm enough to actually open the door. That's a narrow operating window. Luna: So where does that leave us? Are we going to have reliable door-opening robots in the next five years? Lucas: I think we'll see specialized solutions — robots that can open the specific doors in their assigned building. The broader generalization problem is probably a decade away. But the good news is that the research community is finally treating doors as a serious benchmark. There's a new dataset called DoorWorld that includes 500 doors across 10 countries, with full metadata on handle type, force profile, and hinge geometry. That's the kind of resource that accelerates progress. Luna: So the next time I hold a door open for a delivery robot, I'll be a little more understanding. Lucas: Exactly. That robot is facing one of the hardest problems in manipulation. And maybe — just maybe — in a few years, it will be able to return the favor.