Doppler LiDAR and Vision Systems Are Finally Catching Up to What Factory Floors Needed a Decade Ago

I called my old colleague at Siemens last week to ask what he thought about sim-to-real transfer these days. His response was basically that it's gotten a lot better but still requires significant domain randomization and reward shaping to work reliably. Which is exactly what this paper describes doing. So we're not in magic territory yet, but we're in "this might actually be practical" territory.

Look, here's the thing. Both of these papers are research. Neither is a product you can buy. The DPNet team has released their code on GitHub, which is good, but going from research code to something that runs reliably on an industrial platform is, well, it's a journey. I've watched plenty of promising research die in that gap. The quadcopter work is even earlier stage in some ways, though the fact that they got real outdoor flights working is encouraging.

What strikes me about both approaches is that they're tackling the same fundamental problem from different angles: how do you react to things moving unpredictably in your environment? The Doppler LiDAR approach gives you direct velocity measurements. The vision approach infers motion from sequential depth images. Both use learned components to handle the messiness of real-world data. Neither is trying to solve the problem with pure classical control theory, which (and I say this as someone who spent years implementing classical controllers) is probably the right call for highly dynamic scenarios.

I should note that I don't have access to the actual performance numbers beyond what's in the abstracts. The DPNet paper claims superiority over "extensive benchmark schemes" but I haven't dug into what those benchmarks actually are. The drone paper claims successful zero-shot transfer but the outdoor experiments were presumably in controlled conditions. These are limitations worth keeping in mind.

If I were still at Kuka, or at any of the big integrators, I'd be watching this space carefully. The warehouse automation folks especially. Amazon's been pushing the boundaries of what's possible with mobile robots in dynamic environments, and everyone else is trying to catch up. Systems that can actually handle unexpected obstacles, not just stop and wait for a human, would be genuinely valuable.

We're not there yet. But we're closer than we were when I was dodging trash bins in the lab ten years ago. That's something.

Sarah Williams · 4 hours ago · 4 min