Latest / The Robotics Podcast with Fexingo: Autonomous Systems, Industrial Robots, and Hardware / How Robots Are Finally Learning To Adapt
Transcript
- Lucas: There is a specific kind of frustration that haunts every factory manager who has tried to automate anything that isn't perfectly uniform. You know the drill: you have a robot arm programmed to pick up a widget, but if that widget is slightly rotated or has a burr on the edge, the whole operation jams. For decades, the solution was just to build better jigs and fixtures to hold the part in place. Luna: Right, because the alternative was usually just hiring more people to do the fine-tuning by hand, which defeats the whole point of buying expensive hardware in the first place. Lucas: Exactly. But here is the thing changing right now in late 2026. We are seeing a move away from pure position control toward what engineers call compliant control. It is not just about telling the robot where to go; it is about letting it feel its way there. Think of it like the difference between a puppeteer pulling strings and a human hand reaching for a cup. Luna: So instead of brute-forcing the movement, the robot adjusts its grip based on resistance? That sounds like a massive leap in sensor technology. Lucas: It is. And the concrete example I want to look at today comes from a mid-sized semiconductor packaging facility in Arizona. They were trying to automate the insertion of delicate silicon wafers into holders. The old robots would crush one out of every fifty because they couldn't account for microscopic variations in the wafer's thickness. It was costing them nearly two hundred thousand dollars a month in scrap. Luna: Two hundred thousand dollars is a lot of margin gone just because a machine couldn't be gentle enough. What did they actually change in the hardware? Lucas: They swapped out the rigid joints for ones with embedded torque sensors and used a new algorithm that reads force data at a thousand times per second. Basically, the robot feels the contact. If it hits resistance, it backs off microscopically and tries again, rather than forcing the motion. It turns out that adding a few dollars worth of sensors to each joint completely changes the economics of the deployment. Luna: That is interesting because most people think of robot sensors as these bulky, expensive add-ons, not something that fits into a standard joint. Lucas: And that is the misconception we need to clear up. The cost of those sensors has dropped so much that integrating them into the end-effector, the actual gripper part, is now cheaper than the manual labor it replaces. But the software side is where the real magic happens. The robot has to decide when to yield and when to push. Luna: I imagine that decision tree gets complicated fast if you are dealing with objects that have different stiffness levels, like a rubber seal versus a glass lens. Lucas: It does, and that is why this is not just an incremental improvement. It is a fundamental shift in how we design automation cells. In the past, you designed the cell around the robot's strengths. Now, you are designing the robot around the product's variability. It flips the script entirely. Luna: So the robot becomes less of a specialized tool and more of a general-purpose assistant. That sounds like it could solve a lot of the deployment bottlenecks we talked about last month. Lucas: It absolutely does. Last month we discussed the $40 billion gap in adoption, and part of that gap is exactly this rigidity problem. Companies don't deploy robots because they fear the downtime when things go wrong. Adaptive control reduces that fear significantly. Luna: But surely implementing this kind of feedback loop requires a level of programming expertise that most plants don't have on staff. Lucas: That was the barrier six months ago. But the major manufacturers have been pushing proprietary SDKs that abstract the math away. You no longer write code for force vectors. You tell the system 'gently insert' and it handles the calibration. It is becoming a feature, not a custom project. Luna: If the software handles the calibration, then the implementation time drops drastically. That changes the ROI calculation from years to months. Lucas: Precisely. And let us pause here for a second to talk about how we keep this conversation going. If today's deep dive into adaptive control gave you some usable insights for your operations or investment thesis, a couple of dollars a month makes a genuine difference in keeping these episodes produced ad-free. You can support us at buy me a coffee dot com slash fexingo — that's F. E. X. I. N. G. O. Luna: Yeah, and honestly, hearing that listeners are using these details to make smarter hardware decisions is exactly the fuel we need to keep digging into these technical weeds. Lucas: Thanks for that support. Now, back to the semiconductor case. The real surprise wasn't just the reduction in scrap. It was the speed increase. Because the robot could adapt on the fly, they didn't have to slow down the cycle time to ensure safety margins. Luna: Wait, slowing down for safety is standard practice. How does adapting faster actually allow you to speed up? Lucas: Because the old method relied on static safety buffers. You program the robot to move ten percent slower than its maximum capacity just in case it hits something unexpected. With force feedback, the robot knows instantly if it is hitting something unexpected. It can operate closer to its physical limits because it has immediate corrective action. It is like driving with cruise control versus having your foot on the brake. Luna: That analogy makes it very clear. So the efficiency gain is twofold: less waste and faster cycles. That is a powerful combination for any margin-sensitive industry. Lucas: It is. And this is starting to bleed into other sectors beyond semiconductors. We are seeing pilot programs in food processing, where handling soft produce like strawberries has always been a nightmare for machines. A rigid gripper squashes the fruit. An adaptive one can adjust pressure based on the fruit's firmness. Luna: I never thought about strawberries being a robotics challenge, but it makes sense. Every berry is different. This kind of tech could really unlock automation in agriculture. Lucas: It could. And it is not just about the gripper. It is about the entire kinematic chain. When you introduce compliance, you have to manage the energy storage in those joints. If the robot pushes too hard against a wall, that energy has to go somewhere. Poorly managed, it causes vibration and wear. Luna: So the maintenance profile might actually change. Instead of fixing broken parts, you are monitoring the health of the damping systems. Lucas: Spot on. Predictive maintenance shifts from watching for catastrophic failure to watching for drift in the compliance parameters. It is a more sophisticated model, but it is also more accurate. You catch issues before they cause scrap. Luna: This feels like the next logical step after the simulation gaps we covered a while back. If the robot can learn from its environment in real-time, does that reduce the need for digital twins? Lucas: Not eliminate, but it certainly reduces the fidelity required. You don't need a perfect simulation if the robot can correct for imperfections on the fly. It bridges the gap between the virtual and physical worlds in a way that is much more resilient to noise. Luna: That is a huge implication for startups building robotics solutions. They don't need to simulate every possible variable if their hardware is robust enough to handle the unknowns. Lucas: Exactly. It lowers the barrier to entry for innovative designs. You can experiment with unconventional shapes and materials because the robot isn't fighting you at every turn. It is cooperating with you. Luna: I love that word choice. Cooperating. It really reframes the relationship between human designers and machines. Lucas: It should. We spent forty years making robots that obeyed us blindly. Now we are building robots that understand us. The question for investors and operators alike is whether the initial capex premium is worth the operational flexibility over the next five years. Luna: Given the labor shortages and the rising cost of errors, I suspect the answer is leaning heavily toward yes. Especially in high-value manufacturing. Lucas: Agreed. As we wrap up, I want to leave you with this thought. The next wave of automation won't be about bigger robots or faster arms. It will be about smarter touch. The companies that master this nuanced interaction will have a significant advantage in quality and speed. Luna: A subtle shift, but one that touches almost every part of the production line. I think we will see a lot of pilots testing this in the coming quarter. Lucas: Watch for the announcements from the major players in Q1. They are likely to roll out these compliant features as standard options, not just premium add-ons. It is moving fast.