Latest / The Robotics Podcast with Fexingo: Autonomous Systems, Industrial Robots, and Hardware / Why Robot Arms Still Can't Handle a Wiring Harness
Transcript
- Lucas: So you walk into a modern car factory, right? Spot-welding robots everywhere, doors being installed with millimetre precision. And then you see a row of workers performing what looks like a simple task: plugging a wiring harness into a connector. Luna: I've seen that. It's almost jarring — like, this is the same factory where a robot paints the entire body shell. Why can't a robot just pick up a cable bundle and snap it in place? Lucas: That's the question we're digging into today. And the answer comes down to something deceptively simple: a wiring harness is not a rigid object. It's a bundle of cables that bends, twists, snags, and changes shape the moment you touch it. Luna: Right — so it's not like grabbing a steel bracket where the geometry is known. The harness deforms under its own weight. Lucas: Exactly. And that deformation makes it nearly impossible for a traditional robot gripper to plan a grasp. The robot might try to pick it up at one point, but the harness flops, and the connector dangles somewhere else. Every attempt is slightly different. Luna: Is this a sensing problem, or a control problem? I mean, if we gave the robot better cameras and force feedback, could it adapt in real time? Lucas: It's both. Let's talk sensing first. A standard camera sees the harness as a tangle of similar-coloured wires. Depth sensing helps, but the strands are thin and often occlude each other. Researchers at MIT's CSAIL published a paper last year — they used a special algorithm to track the cable's pose, and even then, their robot could only successfully thread a cable through a guide ring about 40 percent of the time. Luna: Forty percent. That's not production-ready. And that's in a lab with controlled lighting, not on a factory floor. Lucas: Right. On the control side, the challenge is that the harness has many degrees of freedom. For a rigid object, the robot can compute exactly where each point will be after a movement. For a cable, you'd need to model every bend and interaction — that's computationally insane in real time. Luna: So what are companies doing instead? Are they just stuck with manual assembly? Lucas: Partly. But there are workarounds. Some manufacturers use passive fixtures that hold the harness in a fixed shape before the robot picks it up. Others use human-robot collaboration: the robot feeds the cable, the human connects it. But the holy grail — a robot that can take a loose harness from a bin and plug it in — remains elusive. Luna: I've heard about tactile sensors being a big deal for this. Like, the robot 'feels' its way through the insertion. Lucas: Yes, they're promising. A company called RightHand Robotics has a gripper with tactile pads that can sense slippage. In tests, it improved cable insertion success from around 50% to 80%. But 80% still means one in five failures, and in a plant running thousands of harnesses, that's not acceptable. Luna: So we're not there yet. What about learning — can a robot train itself to handle cables better? Lucas: Reinforcement learning is a hot area. There's a 2026 study from Stanford where a robot learned to plug in a USB cable after about 10,000 episodes in simulation. But sim to real transfer is tough — the real friction and flexibility never match the simulation exactly. Luna: That's a lot of compute for a single task. And if the cable type changes, you'd have to start over? Lucas: Pretty much. That's the bottleneck — generalisation. A robot trained on a particular harness model won't work on the next one. So we end up with a long tail of edge cases. Luna: So we're still a ways off from fully automated wiring. And that kind of research takes funding. Lucas: Absolutely. If today's tech conversation gave you something usable — a new perspective on why your car's wiring isn't robot-built yet — you can support the show at buy me a coffee dot com slash fexingo. Every bit helps keep these deep dives ad-free. Luna: It's a small gesture, but it genuinely helps. And speaking of small gestures, one thing that helps robots with cables is better simulation — companies like NVIDIA are building digital twins where robots can practice thousands of cable insertions without wearing out hardware. Lucas: That's a great point. Simulation lets us brute-force the learning. But even there, the physical accuracy is a challenge. The cable's stiffness, the friction — if the sim is off by 10%, the policy might fail in the real world. Luna: So maybe the answer isn't one magic sensor or algorithm, but a combination of better tactile feedback, more robust simulation, and smarter fixture design. Lucas: Exactly. And we're seeing progress. In June 2026, Fanuc demonstrated a prototype cell that could handle three different harness types with a 90% success rate — still not perfect, but moving the needle. It uses a multi-fingered hand with force sensing and a vision system that predicts cable behaviour. Luna: Three harness types. That's promising, but when you consider a typical car has dozens of unique harnesses, we've got a long road ahead. Lucas: And that's the takeaway: wiring harnesses remain one of the most stubborn problems in robotics. They're a reminder that sometimes the hardest thing to automate is what humans do without thinking — because our hands and eyes are incredibly good at dealing with floppy, variable objects. Luna: So next time you plug in a cable at your desk, spare a thought for the robot that wishes it could do that. Lucas: Ha! Exactly. Thanks for listening, and we'll be back with another robotics puzzle soon.