Latest / The Robotics Podcast with Fexingo: Autonomous Systems, Industrial Robots, and Hardware / The Warehouse Robot Hand That Finally Picks Anything
Transcript
- Lucas: So there's this stat that's been bouncing around my head since last week: a robotic arm at a logistics hub in Memphis hit a 97 percent pick success rate on mixed totes. That's not a curated bin of identical widgets. That's random stuff — a bag of oranges, a box of cables, a bottle of detergent. Luna: Wait, 97 percent? That's dangerously close to human-level. A few years ago we were celebrating 85 percent on a good day. Lucas: Exactly. And the reason it matters isn't just the number, it's what makes it possible. It's not a new arm, it's a new hand. A gripper that actually feels what it's touching. That's the piece we haven't talked about on this show yet. Luna: We've done a lot of episodes on why arms are clumsy, but the gripper is where the rubber meets the road, literally. Lucas: Right. So let's start with the problem. Warehouses are the wild west for robots. You've got totes arriving with things that vary in shape, weight, stiffness, and slipperiness. A suction cup works great on a flat cardboard box, but it fails on a mesh bag of lemons. Parallel jaws pinch a rigid cube fine, but a floppy plastic bag just slides out. Luna: And that's why most warehouses still have humans doing the final pick. The robot can bring the tote to them, but the actual grabbing is all manual. Lucas: For now. The new approach is to give the gripper a dense array of tactile sensors — essentially a skin. Each square millimeter has pressure sensors that register contact in real time. And then you train a neural network on millions of grasps, so the system learns not just what a bag of oranges feels like, but how to adjust grip force mid-lift. Luna: So it's like teaching a robot to have a sense of touch, not just eyes. Lucas: And touch is the missing sense. Vision tells you where the object is, but it doesn't tell you how much force is needed to hold it without crushing or dropping. That's the tactile feedback loop. Luna: You know, it reminds me of the old research where they had robots try to pick up a raw egg without breaking it. It was all about force control, but they were using rigid force sensors at the wrist, not distributed pressure. Lucas: And that's the key difference. A wrist force sensor measures the total force, but it can't tell you if the pressure is concentrated on one edge, which is what cracks an egg or crushes a tomato. A tactile skin gives you a pressure map, so the controller can adjust the posture of the fingers to spread the load. Luna: That's a really elegant way to put it. So the Memphis pilot—was that using a specific commercial product or a research prototype? Lucas: It was a retrofit on a standard six-axis arm, but the gripper was from a startup called Tactile Robotics, which has been working on this for about six years. They're not household names yet, but their tech is in a few Fortune 500 pilots. The arm itself was a common industrial model, nothing exotic. Luna: That's the interesting part—you don't need a new robot, you just need a smarter end effector. That could make it way more accessible for smaller warehouses. Lucas: Exactly. And that's where the economics get interesting. A retrofitted gripper costs maybe 20 percent of a brand-new collaborative robot system. So you can upgrade existing arms that are already deployed on line loading or palletizing, and suddenly they can do pick and place too. Luna: So the ROI case is a lot faster. I saw a report that labor costs in warehousing have gone up 18 percent year-over-year since 2023, and turnover is still brutal. Automation that can handle the messy stuff becomes a necessity, not a luxury. Lucas: And that's the backdrop for why this matters. We're not just talking about a cool lab trick. We're talking about a tool that can actually replace a human in a role that is physically taxing and repetitive, but previously thought impossible for machines. Luna: It's funny, we've done a whole series on things robot arms can't do, and here we are talking about something they can finally do. Maybe we should do a follow-up episode on what they still can't do even with tactile sensing. Lucas: Oh, there's plenty. Give it a floppy object like a shirt, and even the best tactile gripper struggles. But the fact that we can now handle a bag of oranges is a huge milestone. Luna: You know, this whole conversation about hardware breakthroughs makes me appreciate the work that goes into keeping a show like this ad-free. It's a deliberate choice, and it means we can cover niche topics like grippers without worrying about sponsor demands. Lucas: Yeah, we've never run ads, and that's something we're proud of. It lets us go deep on whatever we think is important, even if it's not the most advertiser-friendly subject. If you value that independence, you can support us at buy me a coffee dot com slash fexingo. It's a simple way to keep the lights on. Luna: And it genuinely helps us keep doing this. Every little bit counts. Lucas: Anyway, back to those oranges. One of the challenges with a bag of oranges is that it's not just the shape, it's the weight distribution. The bag is flexible, so the center of mass shifts as you lift it. A rigid gripper would squeeze too hard in one spot and pop the bag. Luna: So the tactile skin lets the gripper constantly adjust the pressure across the fingers, almost like a human hand does when it's holding something fragile. Lucas: Precisely. And the neural network has learned from thousands of grasps that when you feel a certain pressure pattern, you need to redistribute the force. It's not just a lookup table; it's a learned policy. Luna: What about the hardware itself? Is there a standard for tactile sensors, or is everyone doing their own thing? Lucas: No standard yet. Some use capacitive sensors, some use piezoresistive, some use optical. The startup I mentioned uses a proprietary capacitive array that they say can detect forces as small as a gram. But the real challenge is durability — you're touching sharp edges, abrasive cardboard, sticky residues. The sensor needs to survive millions of cycles. Luna: That's a big barrier to adoption. If the sensor wears out after a few months, the ROI goes out the window. Lucas: Exactly. And that's why we're not seeing this everywhere yet. The tech works in pilots, but the reliability data is still being gathered. The Memphis pilot has been running for six months without a major failure, which is promising, but it's one site. Luna: So what's the next frontier? If tactile sensing solves the 'grab anything' problem, what's left? Lucas: The next challenge is in-hand manipulation. Once you've grabbed the object, can you rotate it, reorient it, and place it precisely? That's where a lot of the current research is going. Think about assembling a phone — you don't just pick up a component, you have to twist it into place. Luna: Right, and that requires even finer control of the fingertips, almost like a human doing a coin roll. Lucas: Exactly. And there are labs that are already demonstrating that with multi-fingered hands, but it's still slow and computation-heavy. But if you look at the trajectory — from simple pick and place to tactile-based picking — you can see where this is heading. Luna: It feels like we're at a inflection point. The hardware is finally catching up to the software. Lucas: It is. And the interesting thing is that the software is becoming the bigger differentiator. The sensor hardware is getting commoditized, but the neural networks that process the tactile data — that's where the proprietary value lies. Luna: So for a startup like Tactile Robotics, their moat isn't the sensor itself, it's the dataset and the trained model. Lucas: Bingo. They've spent years collecting grasps and training their models, and that's not easy to replicate. It's a classic data network effect: the more they use their grippers in the field, the better the models get, which attracts more customers. Luna: That's a really strong position. And it also means that the big robot manufacturers might have to partner or acquire rather than build from scratch. Lucas: We're already seeing some of that. Major robot arms companies are starting to offer tactile grippers as an option, either through their own development or through partnerships. But the market is still young, and there's room for many players. Luna: So what should a logistics manager listening to this take away? Is now the time to invest in tactile grippers? Lucas: I'd say if you have a high-mix, low-volume operation, it's worth running a pilot. The cost is not trivial, but the labor savings can be substantial. And the technology is only going to get cheaper and better. Waiting a few years is also a valid strategy, but you risk falling behind competitors who get the data advantage. Luna: It's a classic innovator's dilemma. Do you jump in early or wait for the tech to mature? Lucas: And of course, we're talking about the technology, but the human element is still crucial. These systems need supervision, maintenance, and continuous improvement. The robot isn't replacing the worker; it's changing the job. Luna: That's a good note to end on. The robots might be getting better at feeling, but they still need people who can think.