Why Welding Bots Fail to Stick: Industrial AI Needs a Bug-Tracking Mindset

The Gap Between Demo and Deployment

Every few months, a new robotics demo goes viral. A humanoid picks up a box, a robotic arm folds laundry, a quadruped sprints across uneven terrain. The comments fill with wonder. Then the video ends, and the robot goes back to its lab. The same story repeats across the industry: impressive capability in controlled settings, but almost no sustained value on a real factory floor.

Wu Zheming, co-founder and CEO of Qunqing Intelligent, has a blunt take on this disconnect. He's spent years trying to make welding robots work in actual production environments, not just in showcases. His company's approach treats the robot not as a standalone machine but as part of a physical AI system, one that requires the same kind of rigorous feedback loops you'd find in software development. In his upcoming talk at AICon Shenzhen, he'll explain why most embodied AI projects fail to cross the chasm from demo to deployment—and what that has to do with bug tracking.

Robots Are Software, but Heavier

Think about how a typical software team ships a feature. You write code, run tests, deploy, monitor, collect logs, fix issues, and iterate. The loop is tight and well understood. Now imagine doing that with a robot arm that costs six figures and can crush a human if misprogrammed. The stakes change, but the underlying need doesn't: you still need to observe failures, isolate root causes, and improve the system over time.

That's where most industrial robotics efforts fall short. They focus on the model, the perception stack, or the control algorithm, but they ignore the operational layer. How do you know when a weld is slightly off? How do you capture that data? How do you feed it back into the system so the next weld is better? Wu argues that this closed-loop feedback is the real moat, not the neural network weights.

Why Welding?

Welding is not glamorous. It's hot, dirty, and dangerous. But it's also a multi-billion-dollar industry with a persistent labor shortage and a high tolerance for automation, if the automation actually works. Wu's team chose welding as their beachhead because it's a high-value, high-difficulty task that demands precision and reliability. If you can make a welding robot that consistently outperforms a skilled human, you've proven something fundamental.

But the challenges are brutal. Welding environments are noisy, non-standard, and constantly changing. Parts arrive with slight variations. Lighting shifts. Metal reflects and distorts. A robot that works on Tuesday might fail on Wednesday because the ambient temperature changed. This is the opposite of a clean, controlled lab setting.

The Physical AI Loop

Wu's framework for tackling this is what he calls the physical AI loop: perception, decision, execution, and feedback. Each stage feeds into the next, and the loop must run continuously, not just during training.

• Perception: The robot must see and sense the environment, including the exact position and angle of the weld joint.
• Decision: Based on that perception, the system plans the weld path, speed, and parameters.
• Execution: The robot arm moves with precision, controlling the torch and wire feed.
• Feedback: After the weld, the system evaluates the result—maybe through visual inspection or sensor data—and uses that to adjust the next attempt.

This loop is essentially a bug-tracking cycle for physical actions. Every deviation is a bug. Every suboptimal weld is a ticket. The system needs to log, analyze, and fix these issues in real time, just like a software CI/CD pipeline.

Why Demos Don't Translate

Most research labs and startups build impressive demos because they control the variables. They set up perfect lighting, use identical parts, and script the task. The robot executes the same motion over and over. It looks flawless because nothing changes.

In a real factory, everything changes. The part is slightly warped. The fixture is misaligned. The operator placed the piece 2 millimeters to the left. The robot's model wasn't trained for that. So it fails. And because there's no feedback loop, it keeps failing the same way. No one logs the failure, no one updates the model, no one tweaks the controller. The robot gets shelved, and the factory goes back to human welders.

Wu's point is that the hard part isn't the initial intelligence—it's the continuous adaptation. A robot that can't learn from its mistakes is just an expensive paperweight.

The Real Barrier: Data and Evolution

In software, we have version control, issue trackers, and automated tests. In industrial robotics, we have none of that by default. Wu argues that the core competitive advantage for any industrial AI system is the ability to collect real production data and use it to evolve. Without that, you're just guessing.

Qunqing's approach is to build this infrastructure from day one. Each welding robot in the field generates a stream of data: sensor readings, weld parameters, visual feedback, success or failure flags. That data gets fed back into the system, refining the perception models, adjusting the control policies, and creating a digital memory of what works and what doesn't. Over time, the system gets better at handling edge cases, exactly like a bug-tracking system that accumulates fixes.

From Welding to General Industrial Robots

If this loop works for welding, it can work for other industrial tasks. Wu's vision is to generalize the approach: build a platform that can be applied to any robotic manipulation task, from assembly to inspection to material handling. The key is not the robot or the model, but the feedback infrastructure that makes continuous improvement possible.

This is a fundamentally different mindset from traditional automation. Classic industrial robots are programmed to do one thing, over and over. They're precise but brittle. If something changes, they break. The physical AI approach, in contrast, treats the robot as a learning system that can adapt to variation and uncertainty. That's the only way to handle the messiness of real production.

AICon Shenzhen: Where the Rubber Meets the Road

Wu will be speaking at AICon Shenzhen on August 21-22, sharing these insights and more. The conference has a packed agenda covering agent engineering, large model infrastructure, AI-native development, and embodied intelligence. But Wu's talk stands out because it addresses the ugliest, most practical problem in AI: making it work reliably in the real world.

He'll dig into why embodied intelligence demos keep breaking new ground while industrial adoption lags, what distinguishes industrial embodied AI from traditional automation, and how to build a physical AI loop that can handle non-standard, small-batch, constantly changing manufacturing environments. He'll also discuss why real production data and continuous evolution are the ultimate moats for industrial agents.

If you're building AI systems—whether for robots, software, or anything else—the lesson is the same: demos are easy, reliability is hard. The only way to get there is to embrace the feedback loop. Track your bugs. Learn from them. And never stop iterating.