Latest / The Robotics Podcast with Fexingo: Autonomous Systems, Industrial Robots, and Hardware / Why Robot Arms Still Can't Open a Door
Transcript
- Lucas: You walk through dozens of doors every day without thinking about it. Pull, push, turn, twist — it's automatic. For a robot arm, it's one of the hardest things we've asked it to do. Luna: Really? Opening a door? I get that folding a T-shirt is hard — we did that episode — but a door handle seems pretty predictable. Lucas: That's what I thought too, until I looked at the data. A 2025 study from MIT CSAIL tested several state of the art robot arms — including the Franka Emika Panda and the KUKA LBR iiwa — on standard hinged doors. They failed about 40 percent of the time. Luna: Forty percent? That's worse than a toddler. Lucas: Worse in some ways. Toddlers have tactile feedback and years of practice. Robots have torque sensors and vision systems, but they still struggle with the basic physics of a door handle. Luna: Before we get too deep — a quick thing. This show stays ad-free because of listeners like you. A couple of dollars a month is genuinely what keeps these going — buy me a coffee dot com slash fexingo, if you've gotten something out of them. Lucas: Yeah, it really does make a difference. Alright — so why are doors so hard? Let's start with the handle itself. Luna: There are basically two types: levers and knobs. Which one gives robots more trouble? Lucas: Knobs, by a long shot. A lever handle is essentially a torque arm — you push down, and the latch retracts. A knob requires a rotational grip plus wrist rotation plus axial pull. It's a compound motion. Luna: So it's not just grabbing and turning. It's coordinating three different force vectors at once. Lucas: Exactly. The robot has to apply enough torque to overcome the spring in the handle — typically about 0.5 to 1.5 Newton-meters — but not so much that it strips the mechanism or loses grip. And it has to do this while the arm is moving through a complex kinematic chain. Luna: And the door itself moves. Once the latch releases, the door swings open, changing the geometry of the task. Lucas: Right. That's the real kicker. The robot's arm is programmed with a model of the door's expected motion, but real doors have different friction, different hinge stiffness, different floor angles. The MIT study found that even a two-degree misalignment in the hinge axis caused a 15 percent increase in failure rate. Luna: So the simulation doesn't match reality. Classic sim to real problem. Lucas: Classic, but particularly acute here because of the forces involved. You can't just use a compliant gripper and hope for the best — you need active force control. Some groups are using impedance control, where the arm behaves like a spring mass damper system that can adjust its stiffness in real time. Luna: That sounds computationally heavy. How fast does the control loop need to run? Lucas: Ideally, at least a kilohertz. The MIT team used a custom controller running at 2 kilohertz with torque feedback from each joint. Even then, the arm would sometimes overshoot the handle or fail to maintain grip during the pull phase. Luna: What about tactile sensing? We've talked about how robot hands lack the surface feedback humans have. Lucas: That's a big part of it. Human fingertips have about 2,500 touch receptors per square centimeter. A typical robot gripper has maybe a dozen pressure sensors. Without dense tactile feedback, the robot can't tell if it's slipping until it's too late. Luna: So they're basically operating blind, force-wise. Lucas: Partly. Vision helps — the robot can see the handle and track the door's motion — but vision is slow and occluded once the gripper contacts the handle. The key is combining vision with torque and tactile data in a hybrid control loop. Luna: Are there any robots that can reliably open doors today? I've seen videos of Boston Dynamics' Spot using its arm to open a door. Lucas: Spot can do it in controlled environments — well-lit, known handle types, pre-mapped door mechanics. But take it to an office building with different handles, different door weights, different thresholds, and the success rate drops. In the MIT test, Spot succeeded about 70 percent of the time on lever handles, but only 30 percent on round knobs. Luna: So we're not at the point where a robot could navigate a building autonomously and open every door it encounters. Lucas: Not even close. That's why most mobile manipulation robots still rely on a human to open doors for them. It's a major bottleneck for deploying robots in warehouses, hospitals, or offices. Luna: It seems like one of those skills that looks easy but turns out to be a research frontier. Lucas: It really is. And there's progress — new torque-sensing joints, better tactile skins, and simulation environments that randomize door parameters more aggressively. But it's a reminder that the physical world is full of variations we don't even notice. Luna: So what's the next step? Are there any startups or labs focused specifically on door opening? Lucas: A few. One that caught my eye is a spin-out from ETH Zurich called HaptiCore. They're building a gripper with over 100 tactile pixels and integrated force control, specifically aimed at door handles and similar tasks. They claim 95 percent success on lever handles in their lab, but they haven't published independent results yet. Luna: That's promising, but I'll believe it when I see a third-party test. Lucas: Fair. The MIT team is also working on a learning-based approach where the robot practices on a simulated door with randomized parameters, then transfers to real doors. Their latest preprint shows about 80 percent success after a few dozen real-world trials. Luna: So we're moving in the right direction, but it's slow. Lucas: Slow and incremental. That's robotics. Every time you think a task is 'solved,' the real world throws a new variable at you. Luna: Like a door with a push plate instead of a handle. Lucas: Exactly. Push plates, sliding doors, revolving doors — each one requires a different strategy. The dream of a general-purpose robot that can navigate any building is still a ways off. Luna: It almost makes you appreciate how much your brain does without you even thinking about it. Lucas: That's the humbling lesson of robotics. The things we find easiest are often the hardest to engineer.