Latest / The Robotics Podcast with Fexingo: Autonomous Systems, Industrial Robots, and Hardware / Why Robot Arms Still Cant Pour a Perfect Cup of Coffee
Transcript
- Lucas: So, Luna, I've been thinking about something that sounds trivial but is actually a massive headache in robotics. Luna: Let me guess — something to do with gripping or grasping? That's usually where the fun starts. Lucas: Close, but it's about pouring. Specifically, pouring a liquid from one container to another without spilling. Like, why can't a robot arm pour a perfect cup of coffee? Luna: Wait, seriously? We have robots assembling car engines and performing surgery, but pouring a drink is the bottleneck? Lucas: Exactly. And it's not just coffee — it's any liquid. Chemical mixing, bartending, even lab automation. The problem is that liquids are incredibly complex dynamic systems. They slosh, they splash, they stick to surfaces. And most robot arms have absolutely no sense of that. Luna: So they grip the carafe with their usual precision, tilt it at a steady rate, and expect the liquid to behave like a solid? That's the naive approach. Lucas: Right. And it fails spectacularly. In 2019, researchers at the University of California, Berkeley set up a benchmark called the 'liquid handling benchmark.' They had a robot arm try to pour water from a pitcher into a cup — nothing fancy. The success rate was something like 20 percent on the first try. And the failures weren't just spills — they were catastrophic overpours, glugs, air bubbles causing erratic flow. Luna: Twenty percent? That's worse than my first attempt at latte art. So what's the core physics problem here? Lucas: It's the meniscus. When you tilt a container, the liquid surface forms a curved shape due to surface tension and adhesion to the container wall. That curve changes dynamically as the liquid level drops. The flow rate isn't linear, it's chaotic. And if the robot tilts too fast, the liquid can separate from the spout and dribble down the side. Too slow, and you get a dribble anyway because surface tension pulls the liquid back. Luna: So you need to control the tilt angle in real-time based on the liquid's behavior. But a standard robot arm doesn't 'feel' the liquid — it's blind to what's happening. Lucas: That's the sensor problem. Most industrial grippers are rigid — they have force-torque sensors at the wrist, but that only tells you the net force on the entire end effector. They can't sense the subtle pressure changes inside a container as the liquid shifts. So researchers have been experimenting with adding accelerometers to the container itself, or even using cameras to watch the liquid surface. Luna: But cameras have latency, right? By the time the robot 'sees' the liquid slosh, it's already sloshed. Lucas: Exactly. And that's where machine learning comes in. A group at MIT, around 2020, trained a neural network to predict the liquid's motion based on the robot's own actions. They used a model predictive control approach — the robot anticipates the slosh and adjusts the tilt trajectory before it happens. They got pouring success rates up to about 80 percent. Luna: Eighty percent is decent, but still not Starbucks quality. What's the remaining twenty percent failure mode? Lucas: It's the little things — the last few drops. When the pitcher is nearly empty, the liquid becomes a thin film. Surface tension dominates. The drops cling to the rim, then fall unpredictably. Some teams have tried using hydrophobic coatings on the pouring lip, but that only helps for some liquids. Others have designed special spouts that direct the flow. But honestly, the 'last drop problem' is an open research question. Luna: So we're basically saying that until robots have a sense of touch as refined as a human fingertip, they'll struggle with liquids. Is there any industrial application that's actually solved this? Lucas: Pharmaceutical labs have commercial systems that handle liquid dispensing, but those use precise syringe pumps, not robot arms. They're essentially automated pipettes. The challenge is specifically with general-purpose robot arms that have to handle arbitrary containers and liquids. In manufacturing, there's a growing need for robots to handle tasks like pouring resin into molds or mixing chemicals — but those environments are highly controlled. The research is really about making robots robust enough to pour a cup of coffee in a messy kitchen. Luna: It's funny — we think of pouring as a 'simple' motor skill, but it's actually a feedback loop involving vision, tactile sensing, and predictive modeling. And we haven't cracked the tactile part yet. Lucas: Right. And interestingly, a lot of the progress is coming from machine learning that simulates the physics of liquids. One team from NVIDIA in 2024 used a digital twin of a pouring task — they simulated millions of pouring attempts in software, each with slightly different parameters. The robot learned a policy in simulation that transferred to the real world with about 90 percent success. That's promising. Luna: I wonder if there's a simpler hack: just use a flexible gripper that conforms to the container shape. Or maybe a soft robot hand that can sense pressure distribution. Lucas: Soft robotics is exactly where a lot of researchers are looking. The problem is that soft grippers — like the ones from Soft Robotics Inc. — are great at grasping but not great at force control. They're pneumatic, so they have a slow response. But there's been work on embedding liquid metal sensors into the gripper pads to measure pressure gradients. That might give the robot a 'skin' that can feel the liquid shifting. Luna: So we might be only a few years away from a robot barista that doesn't spill your latte. But in the meantime, I'll keep pouring my own coffee. Lucas: Yeah, and you know, this actually connects to something broader. A lot of the listeners here are professionals who rely on technology every day. And it's easy to get frustrated when a robot fails at something simple. But the reason we can talk about why robot arms still can't pour a cup of coffee is because of the community that supports shows like this. It's a handful of listeners who chip in monthly through buy me a coffee dot com slash fexingo that literally funds making this many episodes. No ads, no sponsors — just people who find these deep dives useful. Luna: Yeah, and I think that's pretty cool. It means we can keep digging into niche problems like the meniscus issue without having to water down the content. So if you're getting value from the show, that's one way to keep it going. Lucas: Exactly. And now, back to the liquid mess. One approach that's gained traction is using a 'pouring trajectory library' — a database of precomputed tilt profiles for different container shapes and liquid viscosities. The robot recognizes the container and viscosity from a camera and then picks the best trajectory. It's not adaptive, but it works well for known scenarios. Luna: So it's like a recipe book for pouring. But what about when the container is unknown? Like a random wine bottle? Lucas: That's the holy grail. Some labs are working on 'one-shot' pouring, where the robot observes the liquid behavior during the first pour and updates its model for subsequent pours. That's closer to how humans learn — we adjust after the first spill. But it requires real-time perception and fast adaptation, which is still computationally expensive. Luna: I can imagine a future where every restaurant has a robot arm that can pour drinks, but it has to be calibrated for each bottle type. That sounds like a nightmare for scalability. Lucas: It is. And that's why some companies are avoiding the problem altogether. For example, in automated bartending, they use pre-filled dispenser guns rather than robot arms. But for tasks like pouring molten metal in foundries, you can't avoid the arm — and the stakes are higher. A spill there is dangerous. Luna: So for high-temperature liquids, the sensor problem is even worse. Standard electronics can't handle the heat. So they have to use non-contact methods like infrared cameras to monitor the pour. That adds another layer of complexity. Lucas: Exactly. But there's been a breakthrough from a team at the University of Tokyo in 2025. They used a combination of an accelerometer on the ladle and a machine learning model that predicts the liquid's center of mass. They achieved a 95 percent success rate in pouring molten aluminum — within a very controlled setup. It's not ready for prime time, but it's a proof of concept. Luna: So progress is being made, but we're not at the point where you can trust a robot to pour your morning brew without supervision. That's a good reminder that some of the hardest problems in robotics are the ones that seem easiest to humans. Lucas: Absolutely. And this is why I find this field so fascinating. Every time we think we've solved manipulation, we find another edge case. Pouring is just one example — but it encapsulates so many challenges: sensing, modeling, control, and adaptation. Luna: Well, I'm going to pay more attention to my next pour. Thanks Lucas. And for those listening, if you want to support deep dives into why robots still can't do your chores, you know where to find us. Lucas: That's buy me a coffee dot com slash fexingo. And we'll be back next week with another episode. Until then, keep pouring — hopefully without spilling.