Latest / The Robotics Podcast with Fexingo: Autonomous Systems, Industrial Robots, and Hardware / Why Robot Arms Still Can't Open a Door
Transcript
- Lucas: So here's a weird one. You can buy a robot arm today that can weld a car chassis, pick up a 100-kilo crate, or assemble a circuit board at sub-millimeter precision. Luna: But ask it to open a door? Lucas: And it'll probably fail. I'm not talking about a heavy vault door or something with a combination lock. I mean a standard interior door with a lever handle. The kind you walk through twenty times a day without thinking. Luna: And we're not alone in struggling with this. I saw that at the DARPA Robotics Challenge finals a few years ago, several humanoid robots spent minutes just trying to get through a doorway. Some never made it. Lucas: Exactly. That was 2015, and we've made huge strides in locomotion and perception since then. But opening a door remains a benchmark that most research labs still use to test manipulation robustness. Luna: So what makes it so hard? I mean, a door handle is basically a lever. You push down, pull, and the door swings. Why can't a robot just do that? Lucas: Okay, so let's break it down. First, the robot has to find the handle. That's a computer vision problem — detecting a thin metal bar against a flat door surface. Modern systems can do that reliably with depth cameras. But then it has to grasp it. Luna: And grasping is where things get ugly. Lucas: Right. The handle might be slippery, or the robot's gripper might not have the right shape. But even if you get a solid grip, the real challenge starts when you try to turn the handle. Because turning requires applying torque along a specific axis, and the handle's resistance varies. Luna: Varies how? It's just a spring-loaded latch, isn't it? Lucas: It is, but the friction in the latch mechanism depends on how the door is hung, the temperature, the age of the hardware. And you can't model that exactly in advance. So the robot has to feel its way through the motion — that's force control. It has to sense how much torque it's applying and adjust in real time. Luna: So it's a blind, touch-based problem. Vision gets you to the handle, but after that, you're flying by the seat of your pants. Lucas: Exactly. And then, once you've turned the handle, you have to pull or push the door open. But the door's center of mass is moving, the hinge friction changes with the angle, and if there's a draft or a slight slope in the floor, the door might swing differently than expected. Luna: So the robot has to coordinate its arm motion with the door's motion. That's like trying to push a shopping cart while also steering it — but you can't see the wheels. Lucas: That's a great analogy. The robot's arm is essentially coupled to the door through the handle. If it applies too much force, it might yank the handle off or cause the robot to lose its balance. Too little, and the door doesn't open. And the robot can't predict exactly how much force is needed because every door is different. Luna: So how do researchers tackle this? I remember reading a paper from MIT's CSAIL where they used a learning-based approach. They let the robot practice on a few doors, and then it generalized to new ones. Lucas: Right, there's been a lot of work on using reinforcement learning to train policies that map force sensor readings to motor commands. The robot tries, fails, adjusts, and eventually learns a strategy. But there's a catch. Luna: What's that? Lucas: The simulation to reality gap. Most of these policies are trained in simulation because it's faster and safer. But a simulated door doesn't have the exact friction, the exact handle texture, the exact hinge slop. So when you deploy the policy on a real robot, it often stumbles. Luna: So we're back to the hardware problem. Maybe the solution isn't smarter software, but better hardware. Like a gripper that can conform to the handle shape or sense force more precisely. Lucas: That's one angle. There are also specialized end-effectors designed just for door handles. But then your robot loses versatility. The whole point of a general-purpose robot arm is that it can do many things. If you have to swap grippers for every task, you're back to the factory floor. Luna: True. And in a disaster response scenario, you don't get to choose your door. The robot has to handle whatever it finds. Lucas: Exactly. That's why DARPA pushed for humanoid robots that could navigate rubble, turn valves, and open doors. The 2015 challenge showed just how far we were from robust autonomy. And since then, progress has been steady but slow. Luna: Has anyone actually solved it? I mean, are there robots today that can reliably open a random door? Lucas: In controlled lab settings, yes. Several research groups have demonstrated robots that can open a new door after seeing it once. But the success rate drops when you change the environment — different lighting, different handle types, different door weights. Luna: So it's like the grape or the wet paper towel. Another human-easy task that robots can't quite master. Lucas: And that's what makes it fascinating. Each of these failures teaches us something about the gap between biological and robotic manipulation. We have amazing sensors and actuators, but we lack the integrated feedback loops that humans have. Luna: It's also a reminder that real-world robotics is hard. The internet makes it look like robots can do everything, but the edge cases are brutal. Lucas: And that's honestly why we keep doing this show. Because the real story isn't the polished demo video — it's the hours of failed attempts and incremental improvements. Luna: Speaking of which, a quick note. If you find these deep dives into the gritty details useful, consider supporting the show. A couple of dollars a month genuinely makes a difference — it's what keeps the podcast ad-free and independent. Lucas: Yeah, listeners like you are the reason we can spend time digging into papers and talking to researchers instead of chasing ad revenue. If you've gotten something out of the robotics episodes, you can chip in at buy me a coffee dot com slash fexingo. Luna: No pressure, really. Just if you want to help keep this going. Now, back to doors — I want to talk about a specific approach from UC Berkeley that uses tactile sensing. Lucas: Oh, the GelSight-based one? Where they put a camera inside the gripper to map surface contact? Luna: Exactly. They mounted a GelSight sensor on the gripper, which gives them a high-resolution map of contact pressure and slip. The robot can literally see the handle slipping and adjust its grip. Lucas: That's a clever hack. It turns the force-control problem back into a vision problem. And vision is something we're much better at. Luna: Right. In their experiments, the robot could open doors it had never seen before after just a few seconds of exploration. It would wiggle the handle, sense how it moved, then commit to the motion. Lucas: So the key is active perception. Instead of trying to model the door in advance, the robot uses its own actions to probe the environment and build a model on the fly. Luna: Exactly. And that seems to be the emerging consensus in manipulation research. Don't try to plan everything ahead. Use feedback loops to adapt. Lucas: It's the same lesson we've seen with walking robots. They don't precompute every footstep — they use inertial sensors and joint torque to react to the ground. Luna: So maybe the future of door-opening isn't a perfect simulation, but a robot that's really good at feeling its way through. Lucas: I think so. And there's a broader point here. Every time a robot fails at something simple, it forces us to rethink our assumptions about intelligence and control. Doors are a great example because they're everywhere, and they're all subtly different. Luna: Plus, it's a great party trick. Imagine a robot that can open any door. That's one step closer to a truly useful home robot. Lucas: Right. But we're not there yet. And that's okay. For now, I'll keep opening my own doors. Luna: And I'll keep being grateful that I don't have to write the control code for it. Lucas: That's the spirit. Next time, we'll talk about another surprisingly hard task — maybe folding laundry. But that's a whole other nightmare.