Latest / The Robotics Podcast with Fexingo: Autonomous Systems, Industrial Robots, and Hardware / Why Robots Struggle to Open a Door
Transcript
- Lucas: There's this quiet moment at every robotics conference where a humanoid robot stands in front of a door, and everyone holds their breath. Luna: I've seen the videos. The robot reaches out, misses the handle, or it grabs it but then can't figure out whether to push or pull. Lucas: Right. And it's not because the engineers didn't try. Opening a door is a genuinely hard manipulation problem that cuts to the core of how robots understand force, geometry, and constraint. Luna: Harder than packing a suitcase or screwing a lid on tight? Lucas: In some ways, yes. A suitcase lid is a single motion — push down until you hit a hard stop. A door is a sequence: you have to locate the handle, match its orientation, apply force along the release direction, then transition to a sweeping arc motion while maintaining contact. And the handle could be a lever, a knob, a push bar, or a recessed pull. Luna: Every door is a slightly different mechanism. Hotels have those heavy fire doors with push bars. Office doors have lever handles that you have to press down. My apartment door has a round knob you twist. Lucas: Exactly. So the problem is partly perception — the robot needs to classify the handle type and locate its geometry in 3D space — and partly control. Once the robot's gripper contacts the handle, it needs to feel the forces and adjust in real time. If it pushes too hard, it might jam the mechanism. If it pulls too softly, the latch doesn't release. Luna: And most industrial robots are completely blind to that. They're used to rigid, repeatable tasks. A door is a compliant, unpredictable system. Lucas: That's the core insight. A typical industrial arm — say from Fanuc or KUKA — can position a heavy car door with millimeter precision. But it does it by being stiff. It assumes the world is exactly where it expects. A door handle is rarely exactly where the CAD model says it is. Plus, the robot has to deal with the door's own weight, hinges, and any slight misalignment. Luna: So what's the actual physics here? Let's say the robot manages to grasp a lever handle. What happens next? Lucas: The robot has to apply a force vector that's perpendicular to the lever's rotation axis — usually downward — to unlatch it. Then, while maintaining that downward force, it has to pull the door toward itself. The door's hinge creates a circular arc. The robot's wrist has to follow that arc while the base may need to move backward. So it's a coordinated motion between the arm and the mobile base, all while the force at the end effector changes because the door's weight shifts. Luna: And if the robot's base is on wheels, it might skid or roll backward because it's pushing against a fixed object. Humanoids have to shift their stance to avoid tipping. Lucas: Right. There's a famous paper from the Robotics: Science and Systems conference a few years ago where researchers tested a PR2 robot on door opening. They found that the robot failed about 40 percent of the time on the first attempt. The main failure modes were: misclassifying the handle type, losing grip during the unlatch phase, and applying too much force after the latch released, causing the door to swing open faster than the robot could react. Luna: So what's changed since then? Have robots gotten better? Lucas: Better, but not solved. The biggest improvement has been in force-torque sensing. Modern collaborative robots like the Universal Robots e-Series have built-in torque sensors that let them feel contact forces in real time. That allows something called impedance control — instead of commanding a position, the robot commands a stiffness. It acts like a spring. If it hits something unexpected, it gives way rather than slamming into it. Luna: That's exactly what you need for a door. You want the robot to be compliant enough to follow the door's arc, not fight it. Lucas: Yes. But impedance control is computationally expensive, and it requires accurate models of the robot's own dynamics. A lot of humanoid robots still use high-gain position control because it's simpler and faster for walking. So when they try to open a door, the controller fights itself. The walking controller wants stiff legs; the arm controller wants compliant arms. They conflict. Luna: That's the whole-body control problem. Boston Dynamics has shown Atlas opening doors, but they have a custom control stack that's not widely available. Lucas: And even Atlas sometimes fails. There's a blooper reel from a 2023 demo where Atlas tries to open a glass door with a push bar. It reaches out, pushes the bar, but the door is a sliding door — it needs to slide, not swing. Atlas didn't detect that, so it kept pushing inward, and the door didn't budge. The robot just stood there pressing against a locked slider. Luna: So perception is still a bottleneck. The robot sees a bar and assumes it's a push bar, but it's actually a handle for a sliding door. That's a semantic classification error. Lucas: And that's where the research is going now. Combining depth cameras with tactile sensing. There's a group at MIT that developed a GelSight sensor — a soft, gel-based tactile sensor that can detect texture and shear forces. They mounted it on a robot gripper and trained a neural network to classify handle types by feel after initial contact. So the robot touches the handle, feels whether it's a lever or a knob, and adjusts its strategy. Luna: That's clever. It's like a human's exploratory touch. We don't just look at a handle — we touch it to confirm how it works. Lucas: Precisely. The robot learns a kind of haptic exploration. But that adds cycle time. In a factory, every second counts. A robot that takes ten seconds to open a door is too slow for a production line. For a service robot in a hospital or hotel, maybe that's acceptable. But for industrial automation, the door problem is often solved by simply not having doors — or by using automatic sliding doors with a sensor trigger. Luna: Right — the industry often works around the problem instead of solving it. Lucas: And that's a legitimate engineering choice. But if we want robots to operate in human environments — homes, offices, warehouses retrofitted for human use — they'll have to open doors. There's no workaround when the fire code requires a door with a push bar. Luna: Speaking of real-world environments, something that keeps this podcast ad-free and listener-supported is the small but dedicated group of people who chip in monthly at buy me a coffee dot com slash fexingo. It's a modest thing, but it's literally what lets us drill into these niche engineering problems without trying to sell you a CRM. Lucas: Yeah, and we don't want to interrupt the flow too much — but we do record these for people who genuinely care about robotics, and knowing that some of you find value in that keeps us going. So thank you. Now, back to doors. Luna: So if we're talking about commercial applications, which robots are currently best at opening doors? Lucas: The most reliable commercial solution I've seen is from a company called DoorDash Robotics — no relation to the food delivery service — they build autonomous carts for hospitals that can call an elevator and open a door using a specialized end effector. It's a custom gripper with a motorized roller that can turn a knob or push a bar. But it's a purpose-built tool, not a general-purpose hand. Luna: So specialized hardware can sidestep the control problem. If you design the gripper for the task, you don't need a fancy force-torque algorithm. Lucas: Exactly. And that's the pragmatic path for most near-term deployments. But the holy grail is still a general-purpose humanoid that can walk up to any door and open it. That's what companies like Tesla Optimus, Figure AI, and Apptronik are aiming for. And they're making progress, but the door remains a benchmark. The DARPA Robotics Challenge in 2015 had a door-opening task, and many robots failed spectacularly. Luna: I remember one robot — I think it was from Team RoboSimian — just fell over when it tried to push a door. It leaned in, the door didn't move, and the robot's balance controller couldn't recover. Lucas: That's a classic failure. The robot applies force, but the ground reaction force shifts its center of mass. If the ZMP — zero moment point — moves outside the support polygon, the robot falls. So door opening is also a locomotion problem. The robot has to brace its body, possibly step backward as the door opens, all while keeping its torso oriented correctly. Luna: And that's why so many research labs still use wheeled bases for door opening. That's the TurtleBot approach. The base can roll backward naturally, and there's no balance issue. Lucas: Right. The Fetch robot, which is wheeled, can open doors fairly reliably. It has a compliant arm and a laser scanner for localization. But a wheeled robot can't climb stairs, so it's limited to single-floor environments. Luna: So the trade-off is between mobility and manipulation. Legs give you stairs, but they make doors harder. Wheels handle doors, but not stairs. Lucas: That's the fundamental tension. And I think it's why we'll see hybrid designs in the future — maybe a wheeled base that can deploy small legs for stairs, or a humanoid that uses its arms as crutches to stabilize while opening a door. Luna: There's also the possibility of robot swarms — small robots that coordinate to open a door. One robot pushes the handle, another holds the door, a third drives through. Lucas: That is being researched. But coordinating timing and communication adds another layer of complexity. For now, single-robot door opening is still a research problem with no universal solution. Luna: So what's the one thing our listeners should take away today? Lucas: That a door is not just a door. It's a test of a robot's entire sensorimotor pipeline: perception, classification, force control, whole-body coordination, and adaptation to variation. Solve the door, and you've probably solved a lot of other manipulation problems too. That's why every robotics lab still cares about it. Luna: And maybe the next time you push open a door effortlessly, you'll appreciate the complexity. Lucas: I know I will. I already do.