Latest / The Robotics Podcast with Fexingo: Autonomous Systems, Industrial Robots, and Hardware / Why Robot Arms Still Can't Handle a Rubik's Cube
Transcript
- Lucas: So I was in the lab last week watching a six-axis industrial arm try to rotate one face of a Rubik's cube. It took thirty-seven minutes and it dropped the cube on the floor twice. Luna: Thirty-seven minutes? That's worse than I was doing as a kid. And I was never fast. Lucas: Right. And that's kind of the point. A Rubik's cube looks like a simple toy, but for a robot it's an absolute nightmare. Tight tolerances, glossy surfaces, no two faces are identical in friction, and the state changes every time you move a piece. Luna: So we're talking about why robot arms still can't handle a Rubik's cube — and what that reveals about the state of manipulation. Lucas: Exactly. Let's start with the geometry. A standard three-by-three cube has gaps between pieces of about one to two millimeters. That's within the repeatability of many industrial arms — a Fanuc or a KUKA can hit a position within plus or minus 0.05 millimeters. But that's just positioning. Luna: The problem is that the cube isn't rigidly fixed. You're not just moving to a coordinate. You have to apply the right torque to rotate a face without pushing the whole cube out of alignment. Lucas: And that's where force-torque sensors come in. Most arms have them at the wrist. But the cube's glossy ABS plastic reflects infrared — many optical sensors can't get a good read. And the force signals are tiny. Rotating a face requires about 0.2 to 0.5 newton-meters of torque. That's roughly the force of a nickel coin pressing on your fingertip. Luna: So the sensor noise can easily swamp the signal. You think you're applying a gentle twist and instead you're either slipping or overtightening. Lucas: Exactly. And that's just for a single rotation. Now consider the full solve. The robot needs to know exactly which state the cube is in after every move. That means tracking the colors of all fifty-four stickers. But the cube rotates in your gripper — the state gets lost unless you have external cameras. Luna: So most groups working on this use a fixed camera looking at the cube. But then you have occlusion — the gripper blocks several faces. You need to either reposition the camera or regrip the cube mid-solve. Lucas: And regripping is itself a hard problem. To change the orientation of the cube, the robot has to set it down, pick it up from a different angle, or use a second arm. That's finger gating — the same concept from walking. You release some fingers, move the hand, and regrasp. Luna: I remember we did an episode on why robots can't handle a single sheet of paper. The dexterity requirements are similar. But with the cube, the added constraint is that you can't damage the stickers or the mechanism. Lucas: Right. So the state of the art — what's the best we've seen? In 2018, a team at MIT built a custom gripper with two parallel plates that could solve a cube in about 0.38 seconds. But that was a dedicated mechanism, not a general-purpose robot arm. They literally built a machine that only solves cubes. Luna: And it wasn't dexterous — it used a hard-coded sequence of moves, not real-time sensing. It's like a printer that only prints one document. Lucas: More recently, researchers at the University of California, Berkeley, used a humanoid robot with two seven-degree-of-freedom arms and a custom three-fingered hand. That robot can solve a scrambled cube in about thirty seconds. But here's the catch: the fingers are covered in a special high-friction silicone, and the cube is held in a jig during the solve. Luna: So it's still not end to end. The robot doesn't pick the cube up from a table, shake it, and start solving. The cube is already positioned in a holder. Lucas: Right. The pick is itself a challenge. A standard parallel-jaw gripper — the kind on most factory arms — can't reliably pick up a cube from a flat surface because the jaws need to slide under the cube. The clearance is maybe two millimeters. Most grippers will push the cube across the table instead of lifting it. Luna: So you need suction or a scooping motion. But suction on a glossy surface can fail if there's dust or oil. And scooping requires precise trajectory planning. Lucas: Let's talk about the planning itself. Solving a Rubik's cube algorithmically is trivial for a computer — there are known optimal solvers that find a solution in under twenty moves. But the robot has to execute those moves in the physical world, with compliance and error correction. Luna: So the bottleneck isn't the brain — it's the fingers. The computation is easy. The manipulation is hard. Lucas: Exactly. And that's why this is such a good benchmark. A cube tests every axis of dexterity: precision, force control, tactile sensing, state estimation, regrasping, and robustness to slipping. If your robot can solve a cube from a random pick, it can probably handle a lot of assembly and packaging tasks. Luna: I feel like we should mention the elephant in the room: there's a reason most research labs use a custom cube with a magnetic core or QR codes on each sticker. They're trying to simplify the sensing problem. Lucas: And that's fair for a research prototype. But for the cube to be a real benchmark, it should be off-the-shelf. A standard Rubik's brand cube, with stickers, not tiles. That's the honest test. Luna: So where does that leave us in mid-2026? Are we close to a general-purpose arm that can pick up any cube and solve it in under a minute? Lucas: I don't think so. The best integrated system I've seen is from a lab in Zurich — they use a KUKA LBR iiwa with a Schunk dexterous hand. They can solve a cube about sixty percent of the time, with an average solve time of three minutes. But they fail when the cube is slightly misaligned or when the lighting changes. Luna: And those are exactly the conditions you'd have in a real factory. Slightly different parts, varying light, dust on the surface. Lucas: Right. So the cube problem is really a proxy for the general manipulation problem. If we crack one, we crack the other. And that's why companies like OpenAI and Google DeepMind have funded cube-solving projects — not because they want a robot that can play with toys, but because the skills transfer to everything else. Luna: It's also a great demo for investors. A robot solving a cube looks impressive on video. But the hard part is making it reliable. Lucas: Speaking of which — I want to be honest about something. This show is entirely listener-funded. We don't run ads, we don't have sponsors. A small group of people chip in every month through buy me a coffee dot com slash fexingo, and that literally pays for the research, the recording, and keeping this many episodes going. Luna: And because of that, we can spend a whole episode on a niche topic like this without having to pitch you anything. So if the show is useful to you, that's the way to keep it ad-free. Lucas: Exactly. So back to the cube — I think the next breakthrough won't come from better algorithms. It'll come from better hardware. Specifically, tactile sensors that can feel the cube's rotation in real time, like a human can. Luna: There's some interesting work on GelSight sensors — a small camera pressed against a soft gel that captures texture and slip. But those are still too bulky for a fingertip. Lucas: Right. And even if you had perfect sensing, you still need the actuation. The motors in most robot hands are too big to fit in the finger joints. So you end up with cable-driven mechanisms that have backlash and stretch. Luna: So the ideal robot hand for a cube is probably one with at least three fingers, each with three degrees of freedom, equipped with tactile sensors, and driven by direct-drive motors in the palm. Lucas: That hand doesn't exist yet. But several startups are working on it. I think within five years we'll see a commercial system that can solve a cube reliably. And when that happens, a lot of manual assembly tasks will suddenly become automatable. Luna: So the cube is a canary in the coal mine. When the canary sings, the factory changes. Lucas: Exactly. And that's why I find it so fascinating. A simple plastic puzzle exposes the deepest challenges in robotics. Every time I see a robot drop a cube, I'm reminded of how far we still have to go. Luna: And also how far we've come. Thirty years ago, a robot couldn't even pick up a cube reliably. Now at least some can solve it — just not all the time. Lucas: Right. So the question is: will the next generation of robots make the cube look easy? I think they will. But it's going to take better sensing, better hands, and better integration than anything we have today. Luna: I'll be watching the lab videos. And I'll keep a cube on my desk as a reminder. Lucas: Same here. Thanks for listening.