Latest / The Robotics Podcast with Fexingo: Autonomous Systems, Industrial Robots, and Hardware / Why Robot Hands Still Can't Handle a Deck of Cards
Transcript
- Lucas: I've been thinking about a new frontier in robotics that actually hits close to home for anyone who's played a card game. Robot hands still can't handle a deck of cards. Luna: Wait, like shuffle and deal? Or just pick up a single card? Lucas: Let's start with picking up a single card from a flat surface. That alone is a research-grade problem. A team at MIT's CSAIL lab published a paper last year — 2025 — where they built a custom gripper just to do that one thing, and their success rate was 73 percent. Luna: So nearly a third of the time, it fails. That's wild for something so simple. Lucas: Exactly. And it's not because the robot is clumsy in a human sense. The issue is fundamental to how we build grippers. A playing card is thin — about 0.3 millimeters — flexible, and has very low friction. Standard parallel-jaw grippers can't get purchase because there's no edge to squeeze. Suction cups don't work because the card is porous and light. Even magnetic grippers are useless. Luna: So what did the MIT team actually do? What was the design? Lucas: They used what's called a fin ray gripper — those are the flexible, fish fin like structures that bend around objects. But they embedded 64 individual tactile sensors into the fingers. Each sensor measures pressure in real time. So the gripper can feel when it's touching the card and adjust its grip force dynamically. Luna: That's a lot of sensing. How many sensors does a typical industrial gripper have, like in a factory? Lucas: Zero, typically. Maybe one force-torque sensor at the wrist. But for delicate manipulation, you need that spatial resolution. The MIT gripper's sensors are based on a technology called GelSight — it's a soft elastomer with a camera inside that tracks deformation. So it's literally watching the contact patch change shape. Luna: So the robot is using vision through its fingertips, basically. Lucas: Right. And even with all that, only 73 percent success. The failures came from cards that slipped out because the gripper couldn't maintain a stable pinch, or from cards that stuck together — the classic problem of separating one card from another without picking up two. Luna: I remember from earlier episodes, like the one about wet objects or the raw egg, that friction and adhesion are huge problems. Cards seem like a perfect storm — low friction, thin, and they stick together statically. Lucas: Exactly. And here's the thing that makes this more than a parlor trick. If you can reliably pick up a playing card, you can handle paper money, receipts, envelopes, maybe even thin plastic sheets. That's huge for logistics, for postal sorting, for food packaging where you need to separate slices of deli meat or sheets of dried seaweed. Luna: Right, I hadn't thought of the food angle. A lot of food products are thin and flexible and sticky. So this is a commercial bottleneck. Lucas: It absolutely is. A 2024 report from the Robotics Industries Association estimated that flexible object manipulation is the single biggest unsolved problem in warehouse automation. The global market for automated packaging is around 40 billion dollars, and a lot of that is still done by human hands because robots can't reliably pick up a single piece of paper. Luna: That's a huge labor cost. So what's the next step after the MIT card gripper? Is there a commercial product in the pipeline? Lucas: There's a startup called DexForce, spun out of the same lab, that's trying to commercialize the tactile sensing array. They've filed patents on a version that uses printed piezoelectric sensors instead of cameras — much cheaper to manufacture. Their target is the e-commerce returns market, where workers have to open envelopes and extract a single item without damaging it. Luna: That's a great application. Envelopes are basically thick cards with variable contents. Lucas: Exactly. And if you can solve the envelope problem, you can probably handle a lot of other thin objects. But the playing card remains a benchmark because it's standardized — every card is the same size and weight — so it's a clean test of the sensing and control loop. Luna: Has any team actually succeeded at shuffling a deck? That seems like a whole different level. Lucas: Not that I've seen published. There's a 2023 video from the Ishikawa Lab in Tokyo of a robot doing a single riffle shuffle, but it required a custom magnetic deck with embedded markers. The control was mostly pre-programmed, not reactive. So it's not generalizable. Luna: So we're still a long way from a robot that can deal a hand of poker. Lucas: Years, at least. But the progress in tactile sensing is accelerating. Ten years ago, the idea of 64 individual pressure points on a gripper finger was a lab curiosity. Now it's a prototype that works 73 percent of the time on one of the hardest objects we can throw at it. Luna: Speaking of progress, if you enjoy these deep dives into robotics, listener support is what keeps this show ad-free. A couple of dollars a month genuinely makes a difference — buy me a coffee dot com slash fexingo. Lucas: Yeah, it's a small gesture that helps us keep digging into these stories. Back to the cards — I think the next big milestone will be when a robot can pick up a card from a stack without disturbing the rest. That's the deal motion. Luna: Right, because even if you pick one up cleanly, if you lift it, you might drag the others. Lucas: Exactly. The MIT paper called that 'slip-sheet separation' and it's an active research area. Some labs are using air puffs to separate the top card before gripping. Others are using electrostatic attraction — a charged plate that lifts the card slightly. Luna: That sounds like it could damage sensitive items. But for cards or paper, maybe it's fine. Lucas: Right. So there are trade-offs. The point is, the playing card problem is a perfect microcosm of the broader challenge in robotic manipulation: sensing, control, and material interaction all have to work together. And we're not there yet. Luna: But we're closer than we were five years ago. And that's worth noting. Lucas: Absolutely. And maybe in another five years, we'll have a robot that can beat us at poker. Though probably not on the bluffing.