Latest / The Robotics Podcast with Fexingo: Autonomous Systems, Industrial Robots, and Hardware / Why Robot Arms Still Cant Tie a Knot
Transcript
- Lucas: Think about the last time you watched a robot in a warehouse pick up a box. It’s fast, precise, and utterly confident because cardboard is rigid. But take that same robot and ask it to pick up a folded shirt, a tangled cable, or a piece of fabric, and suddenly you have a multi-million dollar system flailing around like it’s blind. Luna: Right, because fabric doesn’t obey Newton’s laws for rigid bodies. It deforms unpredictably, so your computer vision model can’t just map a static coordinate. Lucas: Exactly. And this is the single biggest bottleneck in expanding robotics beyond structured environments like car assembly lines. We call it the soft robotics problem. Today, I want to look at why manipulating compliant materials is so much harder than it looks, using the example of robotic surgical assistants trying to suture tissue. Luna: Surgery seems like an extreme edge case for warehouse logistics. Why start there? Lucas: Because surgery is where the tolerance for error is zero, which forces the engineering to be perfect. In a hospital, a robotic arm needs to grip a needle, push it through flexible tissue, and tie a knot without tearing anything. The tissue stretches, compresses, and moves with the patient’s heartbeat. It’s not just soft; it’s dynamic and nonlinear. Luna: So the challenge isn’t just holding the object, it’s predicting how it will change shape in real-time as force is applied. Lucas: Precisely. Most industrial robots rely on position control. You tell the arm to go to coordinates X, Y, Z, and it goes there. But if you’re pushing against a soft material, the material gives way. So the arm might move three millimeters further than expected, crushing whatever is underneath. That’s why torque control and force feedback become critical here. Luna: But adding torque sensors adds cost and complexity. Does every gripper need that level of sophistication? Lucas: Not every gripper, but certainly any tool interacting with soft matter. Let’s look at the data from recent trials in textile handling. Warehouse robots designed to fold clothes or sort laundry are struggling because they lack what engineers call tactile resolution. They know where the fabric is visually, but they don’t know how hard they’re squeezing it until it slips or tears. Luna: That sounds like a latency issue. If the sensor data takes too long to process, the arm reacts after the slip has already happened. Lucas: It is partly latency, but it’s also signal to noise ratio. Tactile sensors, especially electronic skin prototypes, generate massive amounts of data. Processing millions of pressure points per second requires serious compute power. And right now, embedding that compute directly into the robot’s hand creates heat and weight issues. Luna: So we’re trading off speed for sensitivity. Is there a middle ground where we use vision to compensate for weak touch? Lucas: Some companies are trying hybrid approaches. For example, using high-speed cameras to track the deformation of the fabric in real-time. But lighting changes, shadows, and occlusion make visual-only solutions unreliable. If a corner of the cloth folds over itself, the camera loses track of the geometry entirely. Luna: That makes sense. Vision fails when things overlap. Touch doesn’t have that problem. A finger knows it’s touching two layers even if it can’t see them. Lucas: Correct. Which brings us to the economic impact. If warehouses can’t reliably handle soft goods, they remain stuck handling only standardized boxes. That limits the types of products that can be fully automated. Think about apparel returns. Currently, humans still do most of the sorting because machines can’t distinguish between a wrinkled jacket and a crumpled pair of pants quickly enough. Luna: And returns processing is one of the highest-cost segments in retail logistics. Automating that would save billions. Lucas: Billions. But only if the robots can actually grasp the items without damaging them. Let’s talk about the materials science side. Traditional rubber grips wear out too fast on rough fabrics. Newer materials like silicone composites with embedded microfluidic channels are showing promise. They change shape slightly under pressure, providing a mechanical analog signal before the digital readout kicks in. Luna: Microfluidics? Like tiny water-filled channels inside the rubber? Lucas: Yes. When you squeeze the gripper, the fluid moves. That movement changes the pressure in the channels, which can be measured instantly. It’s a very low-latency way to detect slip or excessive force. It mimics how human skin works with its underlying blood vessels and nerve endings. Luna: Nature really did it first. But can those systems scale to mass production? Manufacturing microfluidic devices sounds expensive. Lucas: It was, but injection molding techniques have improved. Some startups are now producing these grippers at a fraction of the previous cost. The key is integrating them with standard industrial interfaces so factories don’t have to rewrite their entire software stack. Luna: Software integration is always the hidden killer in hardware adoption. If the new gripper requires a custom driver, no factory manager is going to buy it. Lucas: Spot on. And this ties back to the surgical example. The reason robotic surgery hasn’t replaced human surgeons entirely is partly because the haptic feedback—the feeling of resistance—is still lagging behind human intuition. Surgeons can feel the difference between healthy tissue and tumor by the slight give in the instrument. Robots struggle to replicate that nuance. Luna: So the user interface problem is just as big as the mechanical one. If the surgeon can’t feel what the robot feels, they’re flying blind. Lucas: Exactly. Current systems often use visual cues—color changes on screen—to indicate pressure. But that’s a cognitive load. The surgeon has to look at a graph instead of trusting their hands. We need direct haptic feedback where the resistance in the controller matches the resistance in the tissue. Luna: That requires incredibly low latency again. Any delay between the tissue deforming and the surgeon feeling it could cause injury. Lucas: Millisecond delays matter. We’re talking about sub-ten-millisecond response times for true immersion. Right now, most teleoperation systems sit around twenty to fifty milliseconds. It’s noticeable. It feels disconnected. Luna: Twenty milliseconds might sound small, but in fine motor tasks, it’s an eternity. It’s like trying to thread a needle while wearing boxing gloves. Lucas: Perfect analogy. And this disconnect is why many hospitals are hesitant to adopt fully autonomous robotic assistance. They prefer semi-autonomous systems where the human provides the intent and the robot handles the tremor stabilization. It’s a compromise, but it’s safe. Luna: Safe is good. But does that compromise slow down the innovation curve? Are we missing out on capabilities that pure autonomy could unlock? Lucas: Maybe. Pure autonomy could theoretically work faster because it doesn’t wait for human confirmation. But without reliable touch, that speed is dangerous. I think the industry is realizing that you can’t skip the sensory layer. You have to solve the touch problem before you can solve the speed problem. Luna: So the investment is shifting from faster motors to smarter sensors. Interesting pivot. Lucas: Absolutely. Venture capital is flowing heavily into tactile sensor startups. Not just the big tech giants, but specialized firms focused entirely on electronic skin and compliant actuation. Luna: Are any of these companies public yet? Or is it mostly private equity? Lucas: Mostly private. There are a few publicly traded companies making components, like piezoelectric sensors, but the end to end soft robotics integrators are still private. That suggests the market isn’t mature enough for public valuation yet. The technology is still evolving too rapidly. Luna: Rapid evolution means risk. Investors are betting on a winner-take-all scenario in sensor tech. Lucas: Or maybe a fragmented market where different industries use different sensor standards. Medical uses one type, manufacturing another, and agriculture a third. Standardization is unlikely in the short term. Luna: Which makes interoperability a nightmare for large logistics companies trying to deploy robots across different facilities. Lucas: You’re hitting on the infrastructure problem. If every robot brand has its own proprietary sensor protocol, you can’t easily swap out a broken gripper or upgrade a fleet. It locks you into a vendor ecosystem. Luna: That’s a classic platform war. We’ve seen it with charging ports and smart home devices. Will robotics follow the same path? Lucas: Likely. But the stakes are higher here because physical failure can mean safety hazards. A loose electrical plug is annoying; a loose robotic joint holding a heavy package is dangerous. So trust becomes the primary currency. Luna: Trust in the hardware, trust in the software, and trust in the data coming from the sensors. It’s a holistic requirement. Lucas: It is. And that’s why I’m bullish on companies that prioritize open-source sensor frameworks. Even if they don’t sell the hardware, if they provide the standard for how touch data is communicated, they become the plumbing of the next decade of automation. Luna: The plumbing metaphor works. Nobody notices the pipes until they leak. Same with sensor standards. Lucas: Exactly. Now, before we dive deeper into the sensor fusion algorithms, I want to pause for a moment. We’ve been talking about complex engineering problems that require significant resources to solve. Luna: Yeah, building this kind of technology isn’t cheap. Lucas: Right. And honestly, keeping this podcast running without ads is only possible because a dedicated group of listeners supports us monthly. It covers our server costs and lets us keep digging into these technical deep dives without corporate influence. If you find value in these conversations, consider joining us at buy me a coffee dot com slash fexingo. Luna: It really helps us stay independent and focus on the details rather than chasing clicks. Lucas: Thanks to everyone who already contributes. It allows us to keep this show ad-free and focused on substance. Anyway, let’s get back to the sensors. Luna: Where were we? Oh, right, sensor fusion and algorithmic trust. Lucas: Sensor fusion. This is where you combine data from vision, touch, and sometimes sound to create a unified model of the object. The challenge is weighting the inputs correctly. If the camera says the fabric is flat, but the touch sensor says it’s bunched, which one do you believe? Luna: Touch should probably win in that case. Vision can be fooled by perspective. Touch is physical reality. Lucas: Usually yes. But if the touch sensor is dirty or malfunctioning, it might give false positives. So you need a confidence score for each sensor stream. The algorithm has to dynamically adjust its reliance on each input based on reliability metrics. Luna: Like a jury where some witnesses are known liars. You listen to them, but you discount their testimony. Lucas: A great way to put it. And this dynamic weighting is computationally expensive. It requires machine learning models that can run in real-time on edge devices, not in the cloud. Latency is still the enemy. Luna: Edge computing is becoming the standard for robotics. Cloud dependency is a liability for safety-critical systems. Lucas: Agreed. So we’re seeing a trend toward specialized AI chips designed specifically for sensor fusion. Chips that can process visual and tactile data simultaneously with minimal power consumption. Luna: Power efficiency is huge for mobile robots. Battery life dictates operational hours. Lucas: Exactly. A robot that can only work four hours a day because its brain drains the battery isn’t useful in a twenty-four-hour warehouse. Efficiency isn’t just a nice to have; it’s a business requirement. Luna: So the intersection of materials science, sensor hardware, and edge AI is where the real innovation is happening. Lucas: It is. And as those three areas converge, we’ll start seeing robots that can handle the messy, unstructured world outside the factory walls. Delivery bots navigating sidewalks, home assistants folding laundry, surgical teams operating remotely with full tactile fidelity. Luna: It’s a long road, but the pieces are finally starting to fit together. Lucas: They are. And once they do, the definition of 'automation' will expand dramatically. We won’t just be automating repetitive motion; we’ll be automating dexterity. Luna: Dexterity is the final frontier. Once robots can tie a knot, almost anything becomes possible. Lucas: Indeed. Though I suspect tying a knot neatly will still take us a few more years. But the trajectory is clear.