Latest / The Robotics Podcast with Fexingo: Autonomous Systems, Industrial Robots, and Hardware / Why Robot Arms Still Cant Handle a Swinging Pendulum
Transcript
- Lucas: So I was at a robotics demo last week, and someone set up a pendulum — just a metal rod with a weight on the end — swinging it in front of a robot arm. The challenge was for the arm to catch it mid-swing. And the robot just… whiffed, every single time. Luna: Wait — a simple pendulum? Like something a high school student could catch with their hand? Lucas: Exactly. And that's the point. A human can reach out and grab a swinging object almost without thinking. But a robot arm, with all its precision and speed? It's shockingly bad at handling anything that isn't stationary or moving predictably in a straight line. Luna: So what's the fundamental problem? Is it the sensors, the motors, the software? Lucas: All of the above, but let's start with the physics. Imagine a 1-kilogram weight on a 1-meter rod swinging at about 2 meters per second at the bottom of its arc. That's a momentum of 2 kilogram-meters per second. To stop it — or catch it — the robot has to apply an equal and opposite force in a fraction of a second. Luna: That's a lot of torque required almost instantly. Lucas: Right. And most industrial robot arms are designed for precision and stiffness, not for sudden high-torque bursts. They have geared motors and high-inertia linkages. So when the pendulum hits the gripper, there's a delay — the controller has to sense the contact, compute the response, and command the motors. By then, the pendulum has already changed direction. Luna: So the robot is always playing catch-up, literally. Lucas: Exactly. There's a concept in control theory called 'bandwidth' — how fast a system can respond to changes. For a typical industrial arm, the control loop runs at maybe 1 kilohertz, so it samples sensor data every millisecond. But the pendulum's motion changes significantly over 10 to 20 milliseconds. So you're already behind before you start. Luna: Could you use predictive algorithms? Like, estimate the pendulum's trajectory and pre-position the gripper? Lucas: You can, and people do. That's called feedforward control — you model the pendulum's dynamics and send motor commands ahead of time. But here's the catch: the model has to be incredibly accurate. Any friction in the pivot, any slight asymmetry in the weight, any air resistance — and your prediction drifts. After just a few swings, the error is too large. Luna: So the real world is just messy. Models can't capture all the noise. Lucas: Right. And that's why this is such a hard problem. A pendulum is actually one of the simpler dynamic objects — it's a single degree of freedom. Now imagine a crumpled piece of paper tossed in the air, or a water balloon. The degrees of freedom explode. Luna: I remember reading about a robot that could catch a ball thrown by a human. Wasn't that a big deal? Lucas: It was, and it's still impressive. But those systems usually rely on visual tracking and high-speed cameras — they predict the ball's trajectory before it arrives. The catch is, the ball is relatively light and aerodynamically stable. A pendulum is different: it's tethered, so its motion is constrained but also has a restoring force that changes the direction continuously. Plus, it's heavy. Luna: So the pendulum is almost a worst-case scenario for a robot arm? Lucas: It's a great benchmark, actually. Researchers at MIT's CSAIL published a paper last year where they used a special low-inertia arm with direct-drive motors — no gears — to catch a pendulum. They managed it about 60 percent of the time. But the pendulum was only half a kilogram and swinging slowly. Luna: So even with a purpose-built arm, it's not reliable. Lucas: Exactly. And that tells you something fundamental about the gap between biological and robotic manipulation. Human arms have low inertia, high-bandwidth control, and incredibly sensitive tactile feedback. Our muscles can modulate stiffness on the fly. A robot arm is still rigid and slow by comparison. Luna: Okay, but why does this matter practically? Most robots in factories don't need to catch pendulums. Lucas: True, but they do need to handle parts that are moving on a conveyor belt, or tools that are handed to them by humans, or components that shift during assembly. All those scenarios involve dynamic contact — the object is not stationary. If you can't handle a pendulum, you're going to struggle with a lot of real-world tasks. Luna: So it's a proxy for a whole class of manipulation problems. Lucas: Exactly. And the industry is starting to take notice. Companies like Rethink Robotics and Franka Emika have tried to build 'compliant' arms that can yield on contact, but they still have trouble with fast dynamics. The pendulum test is a way to measure how far we've come. Luna: Have there been any breakthroughs recently? Lucas: There's some promising work using learned models — neural networks that predict the pendulum's motion based on past observations, rather than relying on physics equations. That can handle friction and noise better. But it requires lots of training data and still fails when the pendulum's behavior changes abruptly, like if someone taps it. Luna: So we're still a long way from a robot that can play with a pendulum like a kid. Lucas: Yeah, and that's actually exciting. It means there's real engineering to be done. If today's conversation gave you something useful to think about — maybe you're an engineer working on control systems, or just someone who likes understanding why robots are clumsy — I want to mention that this show stays ad-free thanks to listener support. Luna: Yeah, it's a small way to keep the content independent and focused on the tech itself, not on selling ads. Lucas: If you'd like to help, you can find us at buy me a coffee dot com slash fexingo. No pressure, just a way to keep these deep dives coming. Luna: And that support is what lets us spend time on topics like this pendulum problem — which, honestly, is a rabbit hole worth going down. Lucas: So back to the pendulum: one of the most interesting recent developments is the use of 'time-varying' control laws. Instead of trying to catch the pendulum at the bottom of its swing, the robot could intercept it at a point where its velocity is lower — near the apex of the arc. Luna: That makes sense — the speed is zero at the top, so the impact force is minimal. Lucas: Right. But the problem is that the robot has to predict exactly when the pendulum reaches that point, and then move to that location in advance. That's a timing challenge. Humans do it instinctively, but robots need precise state estimation. Luna: So you'd need a high-fidelity sensor to track the pendulum's angle and angular velocity in real time. Lucas: Yes, and that's another bottleneck. Most industrial robots use encoders on the joints, not external vision systems. So they don't even know where the pendulum is unless they touch it. Adding vision introduces latency — even at 60 frames per second, you're looking at 16 milliseconds of delay. Luna: That's like trying to catch something in the dark while looking through a slow-motion camera. Lucas: Exactly. So some labs are using event-based cameras — they only report changes in brightness, not full frames. Those have microsecond latency. Combined with fast control loops, you might get close to human-level performance. Luna: But event cameras are still expensive and not common in industry. Lucas: Right. So this pendulum problem is really a systems problem: you need the right sensors, the right motors, the right control algorithms, and the right materials — all working together. And that's why it's such a good test. It exposes the weakest link in the chain. Luna: It also makes you appreciate how good our own nervous systems are. Lucas: Absolutely. So next time you see a robot arm in a factory, remember: it's incredibly precise at repetitive tasks, but hand it a swinging weight and it'll miss every time. That gap is where a lot of robotics research is focused right now. Luna: And it's a gap that, if closed, could unlock robots in truly dynamic environments — like operating rooms or disaster zones. Lucas: Exactly. So we'll keep watching the pendulum. When a robot can reliably catch one, that's a signal that manipulation has taken a big step forward.