Latest / The Robotics Podcast with Fexingo: Autonomous Systems, Industrial Robots, and Hardware / Why Robot Arms Still Cant Tie a Shoelace
Transcript
- Lucas: Alright Luna, I have a question for you. How long does it take you to tie your shoes in the morning? Luna: I don't know, maybe five or six seconds? It's automatic — I don't even think about it. Lucas: Exactly. That's why it's a perfect example of something that looks easy but is incredibly hard for robots. We've talked about towels, wet soap, slippery fish — but shoelace tying might be the ultimate test of dexterity. Luna: Why is it so hard? It's just a piece of string. Lucas: Right, and that's the deceptively simple part. A shoelace is a flexible, deformable object with essentially infinite degrees of freedom. When you try to grasp it, the shape changes, the forces change, and you need real-time tactile feedback to adjust. Most robot grippers are designed for rigid objects. They either crush the lace or fail to hold it. Luna: So it's not just about the grip — it's also about the sequence of motions, the looping, the pulling through a narrow opening. Lucas: Precisely. Tying a knot requires bimanual coordination: one hand holds the lace taut while the other forms loops, then you pass one loop through a hole and pull. For humans, it's a well-practiced motor skill. For robots, it's a planning problem with contacts that change continuously. Luna: Have any researchers actually succeeded? I remember seeing a video of a robot tying a knot — maybe a surgical suture? Lucas: Yes, there's been work on surgical knots for years. But a shoelace is thinner, floppier, and you have to do it in the real world on your foot. In 2024, a group at MIT's CSAIL published a paper on 'TieBot' — a robot that can tie a shoelace in about 30 seconds. They used a special custom gripper with soft fingers and a camera for visual feedback, combined with reinforcement learning in simulation. Luna: Thirty seconds — that's way slower than my five seconds. But I guess it's a start. How reliable was it? Lucas: The paper reported a success rate of about 60 percent. So it worked slightly more than half the time. Failures included the lace slipping out, the knot being too loose, or the robot getting tangled. Still, it was a big leap from previous attempts which barely managed a single loop. Luna: So the bottleneck is sensing. The robot needs to know where the lace is, how much tension to apply, when the loop is the right size — all in real time. Lucas: Exactly. Vision alone isn't enough because fingers or the lace itself can occlude the camera. That's why tactile sensors — like those from the MIT GelSight or the newer gel-based sensors — are critical. They measure pressure, slip, and shear. Without them, the robot is essentially blind. Luna: I remember we talked about the multimodal sensing approach on a previous episode. But shoelaces must also be different for different materials — some are cotton, some are synthetic, some are round, some are flat. Lucas: That's another layer of complexity. Most research uses a specific type of shoelace. The MIT robot trained on a standard flat cotton lace about 50 centimeters long. Change to a waxed round lace and the dynamics change. So it's not generalizable yet. Luna: What about the application beyond shoes? Why should listeners care about a robot tying a shoelace? Lucas: Great question. This is really about mastering flexible objects. Think about manufacturing: robots that assemble cables, thread wires, or handle fabric. Or think about surgery, where you need to tie sutures precisely. Or even domestic robots that dress people with disabilities. If we can solve shoelace tying, we unlock all of those. Luna: So it's a benchmark, like a cup stacking or a door opening. It's a way to test how close we are to true human-like manipulation. Lucas: Right. And the fact that in 2026 we still have only a 60 percent success rate shows how far we are. But progress is accelerating. Some labs are using soft grippers that conform to the object, like those by Soft Robotics Inc. Others are using parallel-jaw grippers with multiple degrees of freedom. The key may be in combining both. Luna: Where do you see the next breakthrough coming? Better hardware or better software? Lucas: Both, but I'd say software might edge ahead. The hardware — tactile sensors, soft grippers — is already pretty good. The real challenge is the control policy: how to adapt in real time when the knot slips or the lace twists. Reinforcement learning in simulation has let us train behaviors that generalize better, but the sim to real gap remains big. Luna: It's funny — something so trivial to us is still a grand challenge for robots. I'll think about that the next time I'm in a hurry tying my shoes. Lucas: If today's episode gave you something useful — a new way to think about dexterity, or just a neat MIT project — the way we keep this podcast ad-free is through listener support. You can help at buy me a coffee dot com slash fexingo. It takes maybe five seconds, and it directly lets us keep doing deep dives like this. Luna: Yeah, I'll second that. A coffee's worth a bit of knowledge. Lucas: Back to the topic: one idea I find really promising is using modular skills. Instead of trying to tie the whole knot from scratch, you train separate neural networks for 'form loop', 'pass through', and 'pull tight'. Then you chain them with a supervisory system that detects when one skill finishes. Luna: That sounds like a more scalable approach. And you could reuse those skills for other tasks — like tying a bow on a package or closing a bag. Lucas: Exactly. And that's where I think the next two to three years will take us. A modular library of manipulation primitives — for grasping, looping, pulling — that can be combined for any flexible object task. Then tying a shoelace becomes just one application. Luna: What a time to be alive. I'll keep an eye on the MIT lab. Lucas: Definitely. And if you're intrigued by this, check out our earlier episodes on folding a fitted sheet or picking up a wet bar of soap — those tie into the same sensing and control challenges. Next time, we'll look at a new approach to force control that might make all of this easier. Luna: Sounds like a plan. Until then.