How do you build a model that survives degraded or missing sensor data — instead of one that only works in the lab? And the paradox at its center: to learn from bad data, do you still need good labels? （容错训练与标注质量的悖论）
You want a model that handles defective, noisy, or missing sensor data — so it's tempting to think "then I should train on messy data and stop obsessing over data quality."
But the resolution is a clean separation that most people blur: there are two different things that can be "corrupted" — the inputs (sensor readings) and the labels (ground-truth answers). Robustness wants corrupted inputs. It does not want corrupted labels. You deliberately degrade what the model sees while keeping pristine the answer it's graded against.
Hold that distinction — it answers your label question at the end, and it's the through-line of everything below.
Two models perceive the same scene: one trained only on perfect data, one trained with sensor dropout. Turn sensors off and watch them diverge. （关闭传感器，对比两种模型）
The asymmetry is real: studies show fusion models lean hardest on LiDAR. Drop it and a naïvely-trained BEVFusion can lose ~89% of its mAP; drop the camera and it barely flinches. A dropout-trained model degrades gracefully instead of collapsing.
Three layers work together: what you feed the model, how you build it, and how you grade it. Tap each.
Neither real data nor simulation gets there alone. They form a cycle, each fixing the other's blind spot. （真实数据与仿真的闭环）
Real logs capture the true distribution and the weird stuff no one would think to simulate. But failures and edge cases are rare, and labels are expensive.
Use real clouds to calibrate the LiDAR simulator — its noise, dropout, intensity, density — until synthetic and real are statistically indistinguishable. This is the sim-to-real gap, closed deliberately.
Now generate unlimited failure cases with perfect labels for free: every weather, every sensor dropout, every once-a-year edge case — at any severity, on demand.
Train on real data for authenticity plus synthetic corruptions for coverage. The model sees failures it would take a million real miles to encounter.
Test on held-out real corruptions. Where the model still fails tells you what reality to capture or simulate next. The loop tightens.
Real data owns the distribution; simulation owns the coverage. Real tells you what the world actually looks like; sim lets you explore the tail of it safely and cheaply. Robustness comes from the product of the two, not either alone — which is precisely why a strong simulation/world-model capability is the lever for data quality, not a side quest.
The crux. If we're training for messy inputs, can we relax on label quality? Here's the precise answer.
Corrupt the inputs freely. Keep the labels cleaner than ever. The two are independent axes, and conflating them is the mistake.
| Axis | What you do for robustness | Why |
|---|---|---|
| Input quality 输入 | Deliberately degrade | Teaches the model to cope with real-world defects & failures |
| Label quality 标注 | Keep pristine | It's the answer key — noise here is learned as truth, capping accuracy |
1. The signal is harder to find. When inputs are noisy, the model leans even more on the label to know what's real vs. what's a sensor artifact. A wrong label in a noisy cloud is indistinguishable from the noise — it teaches the artifact as the object.
2. Robust training needs a trustworthy compass. Techniques like dropout and corruption-augmentation only work because the loss reliably points "right." Noisy labels break that compass exactly when you're asking the model to do something hard.
3. You can't measure robustness against a noisy ruler. Your evaluation labels must be near-perfect, or resilience-rate and corruption-error numbers are meaningless — you'd be grading degraded inputs against degraded answers.
The requirement shifts even if the quality bar stays high:
▸ Clean-condition ground truth for corrupted inputs. The gold pattern: capture the scene with extra sensors / multi-frame aggregation / a better rig to get a pristine label, then pair it with the single degraded frame the deployed model will actually see. Train input-degraded → label-clean.
▸ Visibility & condition metadata. Each label tagged with occlusion level, weather, sensor health — so you can stratify and weight, not just fit.
▸ Coverage over the failure tail. The new scarcity isn't "more labels," it's "labels for the rare broken cases" — which is exactly what simulation supplies for free with perfect ground truth.
Corrupt the inputs. Never the labels.
Robustness is built by widening what the model can see and survive, while holding the answer key pristine. High-quality labels don't become less important when you train for failure — they become the fixed point that makes training for failure possible. And the cheapest source of perfectly-labeled failure cases is a well-calibrated simulator. （输入可以脏，标签必须干净——而完美标注的失效样本，正是仿真的主场）