Latest / The Robotics Podcast with Fexingo: Autonomous Systems, Industrial Robots, and Hardware / Why Robot Hands Still Can't Handle a Coffee Cup
Transcript
- Lucas: I want to start with an image. Picture a ceramic coffee cup — the kind you'd get at a cafe, no handle, just a smooth cylinder about eight centimeters tall, filled maybe three-quarters full with hot water. Luna: Standard takeaway cup, basically. Lucas: Exactly. Now imagine a robot arm — a Franka Emika Panda, which is a seven-axis collaborative arm that costs about thirty-five thousand dollars — trying to pick that cup up off a table. It has a two-finger parallel gripper with rubberized pads. And it fails. Almost every time. Luna: Fails how? Does it drop it? Tip it over? Lucas: Both, and a third failure mode that's more subtle. A team at MIT's CSAIL published a paper at the IEEE International Conference on Robotics and Automation last year — 2025 — where they ran exactly this test. They set up the cup, the arm, a motion-capture system, and force-torque sensors in the wrist. They ran two hundred and forty trials. The robot succeeded in picking up the cup with a stable grasp only forty-seven times. That's about a twenty percent success rate. Luna: Twenty percent? A toddler can do this by eighteen months. Lucas: Right. And that's exactly the point. The researchers documented three distinct failure categories. The first is what they call 'squeeze-out' — the gripper closes with too much force, the cup deforms slightly, the liquid sloshes, the center of mass shifts, and the cup rotates in the gripper until it slips free. Luna: So it's a force-control problem. The gripper doesn't know how hard to squeeze. Lucas: Partly. But even when they tuned the grip force to a known optimal value — they pre-measured the cup's stiffness and the coefficient of friction — they still got failures. That's the second category: 'tilt and tip.' The arm lifts the cup at a slight angle — less than two degrees — because the wrist's torque sensors have a noise floor of about 0.05 newton-meters. That tiny tilt is enough to let the cup's bottom edge catch on the table surface as it rises, which torques the grasp open. Luna: So the arm can't detect that it's not lifting perfectly vertically. Lucas: Correct. And the third failure mode is the most interesting to me: 'handle hallucination.' The robot's vision system — an Intel RealSense depth camera — sees the cup and plans a grasp at a certain height. But because the cup is reflective and the water inside creates a curved refraction pattern, the depth map at the rim is consistently off by about three millimeters. The gripper ends up contacting the cup one or two millimeters lower than planned, which changes the leverage point and makes the grasp unstable. Luna: Three millimeters. That's tiny. Lucas: It is. But in manipulation, those millimeters are the difference between a stable grasp and a coffee spill. The human hand has about seventeen thousand mechanoreceptors per square centimeter in the fingertips. A two-finger parallel gripper typically has two force sensors total — one per finger, usually a single-axis load cell. The information density is just not comparable. Luna: So what did the MIT team do to improve the success rate? Lucas: They added a passive compliance layer — essentially a soft silicone sleeve with embedded capacitive tactile sensors — over the gripper pads. That brought the success rate up to sixty-two percent. Not great, but a clear improvement. The interesting thing is that the improvement didn't come from better force control. It came from the sensors detecting the onset of slipping before the cup actually moved. Luna: So the robot could feel that the cup was starting to rotate and adjust its grip in real time. Lucas: Exactly. The researchers called it 'pre-slide detection.' The tactile sensors pick up shear strain patterns at the contact surface — something a rigid gripper can't do. The robot then increases grip force by about fifteen percent, and that's enough to prevent the slip from propagating. But there's a trade-off: the soft sleeve deforms, which changes the effective geometry of the grasp, so the position accuracy of the end effector drops by about 0.8 millimeters. Luna: So you trade positional accuracy for tactile sensitivity. That's a real engineering constraint. Lucas: It is. And it's one reason why warehouse automation companies — pick and place for e-commerce — still overwhelmingly use suction grippers for anything that isn't a rigid box. If you have a deformable or liquid-containing object, the entire control paradigm changes. Luna: This also explains why we don't have robot baristas yet, despite all the hype. Lucas: Exactly. A barista's job is not just making coffee — it's picking up cups, wiping counters, opening lids, pouring milk. Each of those actions involves objects that change shape, mass, and friction properties in milliseconds. The MIT paper basically shows that even a state of the art research arm with a custom tactile sensor can't reliably do the simplest version of that task: pick up a cup that isn't moving. Luna: But there are companies like Covariant and RightHand Robotics that claim to have solved general grasping. How do they square with this result? Lucas: Great question. The key distinction is that those companies are doing pick and place in structured environments — bins, conveyors, known item geometries. Covariant's system uses deep learning trained on millions of grasps, but it still fails on novel objects or unusual poses. The MIT test is harder because the cup is handleless, reflective, and contains liquid — three edge cases that don't show up often in warehouse bins. The paper argues that the robotics community has been optimizing for speed and repeatability, not for the kind of fragile, adaptive manipulation that humans do without thinking. Luna: It's humbling. Thirty-five thousand dollars in hardware, years of research, and you still can't reliably pick up my morning latte. Lucas: And that's exactly why this episode matters. The gap between what robots can do in controlled factory settings and what they'd need to do in a cafe or a home is still enormous. But the MIT work also gives a roadmap: better tactile sensing, passive compliance, and — maybe most importantly — algorithms that can handle the uncertainty of objects that change mid-grasp. Luna: So what's next? Is there a version of this that actually works? Lucas: The authors are working on a gripper with four independently controlled fingers, each with a tactile array, combined with a real-time physics model that predicts the liquid slosh. They call it the 'liquid-aware grasp planner.' Early simulations suggest they can get success rates above ninety percent, but they haven't published hardware results yet. That's probably a 2027 paper. Luna: I'll be watching for it. In the meantime, I'm going to appreciate my barista a lot more. Lucas: Same. And actually — if today's conversation gave you something usable, a new way to think about robot dexterity, the way this show stays ad-free is listener support. People who find value in these deep dives can help keep them coming by buying us a coffee at buy me a coffee dot com slash fexingo. Luna: Yeah, it's a small gesture that makes a real difference. We keep the episodes focused and specific because of that support. Lucas: Exactly. And back to the coffee cup — there's one more detail from the paper that stuck with me. The researchers tried using a human demonstration: they had a person pick up the cup while wearing a motion-capture glove, then trained a neural network to reproduce the trajectory. The robot replicated the motion perfectly, but it still failed fifty-three percent of the time. Luna: Because the robot wasn't feeling the cup, it was just following a path. Lucas: Right. The human grasps not by planning a trajectory, but by continuously sensing and adjusting. The robot was executing a script. And that's the fundamental challenge: until robots have sensors that approach human tactile bandwidth, and algorithms that can close the loop at millisecond timescales, tasks like picking up a coffee cup will remain harder than they look.