Latest / The Robotics Podcast with Fexingo: Autonomous Systems, Industrial Robots, and Hardware / Why Robot Manipulation Still Lags Behind Human Touch
Transcript
- Lucas: So we talk a lot on this show about how robots can weld a car chassis in under a minute or pick and place thousands of parts an hour. But there's this quiet, stubborn problem that's been sitting in the corner of robotics research for decades — and it's the reason your Roomba still can't pick up a stray sock. Luna: Right — manipulation. Not just grabbing something, but actually handling it with the kind of dexterity a human hand has. Lucas: Exactly. And I mean really basic things. Like, can a robot arm thread a needle? Can it pick up a raw egg without smashing it? Can it fold a towel? We did an episode on folding towels a while back, but the problem is much deeper than laundry. Luna: Yeah, we covered towel folding in episode 13. But I think the core challenge there is what robotics researchers call 'contact-rich manipulation.' Every time the robot touches the object, the physics change — the forces, the friction — and the robot has to adjust in milliseconds. Lucas: Right. And humans do this without even thinking. You pick up a coffee mug — you don't calculate the grip force; you just feel it. Your fingertips have thousands of mechanoreceptors that measure pressure, shear, texture, and slip in real time. Robots don't have that. Luna: They're basically blind when it comes to touch. Most industrial grippers are just two fingers that close with a fixed force. If the part is a millimeter off, or the surface is a little slick, it either drops it or crushes it. Lucas: And that's for simple pick and place. For something like inserting a plug into a socket — which is a classic assembly task — the tolerances are tiny, and the robot has to wiggle and feel its way in. Humans do that naturally; robots need complex force-torque sensors and iterative algorithms. Luna: There was a great paper from MIT's CSAIL last year — they developed a gel-based tactile sensor that actually measures shear force and contact geometry. It's like a soft fingertip with a camera inside, watching the gel deform. Lucas: Yeah, the GelSight sensor. It's been around in various forms, but they've gotten it small enough and fast enough to use on a real robot hand. The idea is that if you can measure the shape and texture of a surface at high resolution, you can start to do things like detect when an object is slipping and adjust grip in real time. Luna: And that's huge for assembly lines. I was reading about how EV battery pack assembly requires inserting heavy, fragile modules into tight frames — if the robot misjudges even slightly, it can dent the cell or short something. Lucas: Exactly. The automotive industry is a massive driver of this. As carmakers transition to electric vehicles, they're dealing with new materials — soft copper foils, brittle ceramic separators, irregularly shaped battery cells. Traditional rigid grippers don't work well. So companies like Fanuc and ABB are investing in new end-effector designs that incorporate tactile sensing. Luna: Even Boston Dynamics — they're known for locomotion, but their latest research on Atlas, the humanoid robot, shows it can pick up and move heavy objects like car parts. But it still uses a pretty basic parallel gripper. They're working on more dexterous hands. Lucas: The thing is, a human hand has 27 degrees of freedom. Most robot hands have 6 to 12. And even if you add more joints, you need the control software to coordinate them all. That's a computational nightmare — you have to plan a trajectory in a high-dimensional space while dealing with unpredictable contact forces. Luna: So it's both a hardware and a software problem. And the hardware side is getting interesting — there are now 'soft' grippers that use pneumatics or jamming of granular material to conform to objects. But they're still slow and not very precise. Lucas: Right. The jamming gripper — it's basically a balloon filled with coffee grounds. You press it onto an object, suck the air out, and it hardens around the shape. It's great for irregular objects, but it can't do fine manipulation because it's essentially a single 'finger.' Luna: And then there's the whole question of feedback speed. Even if you have a great tactile sensor, the robot's control loop has to run at hundreds of hertz to react quickly enough. That means the sensor data has to be processed and sent to the motors with minimal latency. Lucas: Which is why many research groups are moving toward event-based sensors — they only send data when something changes, rather than streaming a full video at 30 frames per second. That cuts latency dramatically. Luna: So are we close to a breakthrough? I mean, when will a robot be able to, say, peel a banana reliably? Lucas: Funny you ask — that's actually a benchmark task at some labs. Peeling a banana requires the robot to apply precise force to break the skin without crushing the fruit, then grip the peel and pull it off in one piece. As of early 2026, I think the best results show about a 60 percent success rate. Which is impressive, but not ready for your kitchen. Luna: And that's under controlled lab conditions. In a factory with variable lighting, vibrations, and dust, the success rate would drop even further. Lucas: Right. But the progress is real. A few years ago, even a 30 percent success rate on banana peeling was unheard of. The combination of better tactile sensors, more compliant actuation, and reinforcement learning is starting to pay off. Luna: So we're not going to see android hands in factories tomorrow. But maybe in the next five to ten years, some of these technologies will make their way into production lines for tasks that currently require human assembly. Lucas: Yeah, and I think the first big application will be in electronics assembly — inserting cables, connecting tiny ports, handling delicate components. That's where the economic incentive is strongest, because those tasks are still mostly manual and very expensive. Luna: It's also an area where the margins are thin — a single damaged component can cost hundreds of dollars. So even a moderately reliable robot would save a lot of money. Lucas: Absolutely. And that's the kind of concrete use case that justifies the R&D investment. We'll keep watching this space. Luna: Honestly, if today's tech conversation gave you something usable — maybe a new way to think about robot dexterity — that's it. A small gesture at buy me a coffee dot com slash fexingo keeps this show ad-free and focused on the hard problems. Lucas: Yeah, it's a small ask for a big return in our opinion. And it genuinely helps us keep digging into topics like this one. Luna: So, circling back — one more thing I want to mention: there's a startup called Soft Robotics Inc. that's commercializing a gripper with integrated tactile sensing. They claim their gripper can handle fresh produce without bruising. If that works at scale, it could really change food processing. Lucas: I've seen their demos. They use a soft, inflatable finger with a camera inside, similar to the GelSight concept. It's still early, but it shows the direction the industry is heading: away from rigid claws and toward soft, sensing hands. Luna: And that's the kind of shift that could make robots useful in more than just heavy manufacturing. Healthcare, domestic service, even agriculture — all those fields need delicate touch. Lucas: Exactly. So the next time you effortlessly pick up a grape or button your shirt, spare a thought for the roboticists trying to program that. It's harder than it looks. Luna: Thanks for listening to The Robotics Podcast. Lucas: We'll be back next week with another angle on the machines that are reshaping our world.