The Annotation Team worked around the clock for a whole week.

Finally, they finished frame-by-frame precision labeling for all five-hundred-plus selected boundary samples.

The spot-check pass rate reached ninety-one percent, exceeding the eighty-five percent baseline set by the Algorithm Group Leader.

The data augmentation script ran for two nights, generating eight to twelve variations per original sample and expanding the training subset to over six thousand samples.

In the early hours of the eighth day, the second round of training was launched.

This time, he didn't wait around staring at the loss curve.

Because he knew the variable this time wasn't parameters, but data quality.

As long as labeling accuracy improved, whether the model could make the right call wouldn't be known until the run finished.

When Xiao Zheng arrived at his desk the next morning, the training had already completed eight epochs.

"Boss, the curve is out."

The Algorithm Group Leader sat up from his cot. He hadn't gone home last night, working overtime straight through.

On the screen, the loss curve dropped nine percent over the first three epochs.

After that, it began to slow, ultimately leveling off eighteen percent lower than the previous round.

"What about the confusion rate?"

Hearing this, Xiao Zheng pulled up the test set report.

The boundary sample confusion rate had fallen from fourteen percent to nine percent.

"Didn't hit eight."

The Algorithm Group Leader stared at the numbers.

"Off by one point,"

Xiao Zheng said.

"But you have to look at the breakdown.

Multiple people in frame dropped from thirty-one percent to seventeen percent, and soft flooring went from twenty-four percent down to thirteen percent.

What's dragging it down are mainly those completely silent and motionless cases."

The Algorithm Group Leader clicked into the breakdown.

The confusion rate for completely silent and motionless scenarios had dropped from forty-two percent to thirty-five percent.

Although it fell seven points, it remained the highest among all categories.

"How many samples of this type are there in total?"

"Out of the five hundred and thirty-seven labeled samples, sixty-one are completely silent and motionless."

"Sixty-one samples with a thirty-five percent confusion rate means over twenty are still misclassified."

He went through the misclassified samples one by one.

When he reached the seventh, he paused.

On the screen, a three-year-old girl slid from the edge of a coffee table onto the rug, bumped the back of her head, and then went completely still.

No crying, no screaming, not even reaching up to rub the back of her head.

She just sat there, eyes open, staring straight ahead.

However, the shock recorded by the vibration array was 0.29g, well within the threshold range.

The audio spectrum showed no anomalies.

The IMU detected an acceleration change lasting about 0.15 seconds.

Among the three signal channels, both vibration and IMU triggered, but neither amplitude was particularly high.

The joint confidence was evaluated as 'moderately suspicious' and pushed to manual confirmation.

Then, manual confirmation labeled it as a 'true fall'.

Yet the model classified it as 'normal sitting posture' with sixty-one percent confidence.

Xiao Zheng asked,

"Why did it judge it as sitting?"

Hearing this, the Algorithm Group Leader pulled up the model's attention heatmap.

The model was looking at the girl's torso posture: upper body upright, hands resting on her knees, head centered.

From a static posture standpoint, it indeed looked like normal sitting.

The model didn't'see' the impact to the back of her head because that movement lasted less than 0.3 seconds.

Immediately afterward, she settled back into a sitting posture.

"The model is learning the end state,"

the Algorithm Group Leader said.

"It sees the child sitting in the end with hands on knees and head upright, so it classifies it as normal.

That's because it hasn't learned temporal traceback yet; it hasn't learned to look at 'how she ended up in this posture.'"

"The dynamic process?"

"Exactly.

When we labeled, we marked skeletal keypoints for every frame, but the loss function used during model training primarily optimizes for end-state classification.

So the information from the dynamic process got diluted."

Then, he closed the heatmap and opened the training configuration.

"Add a temporal branch.

We don't need to touch the backbone network; we just need to stack an LSTM layer on top, dedicated to processing the temporal features of action sequences.

Make the model look not just at 'what it looks like at the end,' but also 'what happened in between.'"

"Is there enough training data?"

"Among the boundary samples with complete action sequences, there are over six thousand augmented variations. That's plenty.

However, the parameter count of an LSTM is considerably larger than pure visual classification."

He paused and opened the computing power spec sheet.

"The NPU has 4 TOPS of computing power, which is enough to run it.

But the inference latency will increase significantly from the current eighty milliseconds."

Xiao Zheng did some quick math and said,

"Adding an LSTM layer with 128 hidden units and a sequence length of thirty frames adds about thirty milliseconds.

That's roughly one hundred and ten milliseconds in total."

"One hundred and ten milliseconds.

Add the fifty milliseconds for joint vibration and audio inference."

"End-to-end latency will be one hundred and sixty."

Checking against the specifications set at the briefing, the Algorithm Group Leader said,

"The robotic arm touch emergency-stop response time requires under one hundred milliseconds.

However, under the alert tiers, the orange suspected risk has no hard latency requirement; only the red emergency trigger is capped at one hundred milliseconds."

"So the LSTM branch is only used for the orange alert tier?"

"Right.

Red triggers take the lightweight fast path without LSTM, relying purely on vibration plus IMU hard threshold judgments. That keeps latency under seventy milliseconds.

Meanwhile, orange alerts go through full multimodal inference, including visual temporal analysis, allowing one hundred to one hundred and fifty milliseconds."

With that, he began modifying the training configuration, added an LSTM branch, and restarted the training run.

This time, it took fourteen epochs to converge. The loss curve dropped another eleven percent compared to the previous round.

The confusion rate dropped from nine percent down to five percent.

In particular, the confusion rate for completely silent and motionless scenarios fell from thirty-five percent to eighteen percent.

The Algorithm Group Leader looked at this number.

From forty-two percent to eighteen percent—two iterations had cut it by more than half.

However, three out of eighteen samples were still misclassified.

"It's impossible to reach zero,"

he said to Xiao Zheng.

"In this kind of scenario, physical signals are inherently weak.

Even though the child took a bump, the force wasn't large, and she didn't cry or move.

So our vibration sensors and IMU could only catch a faint trace.

The model can recover most cases using temporal analysis, but there will always be edge cases where the signal is too faint to catch anything."

"Then what do we do?"

"Write it into the limitations statement.

On the S02 feature description page, front-load a disclaimer:

'In edge-case scenarios on soft flooring where the subject remains completely motionless and silent, missed detections may occur.

Parents are advised to maintain line of sight when children are playing alone.'"

Having said this, he paused.

"This is also the ceiling of what we can achieve.

What's left isn't an algorithm problem—it's simply that the physical world isn't providing enough information."

Xiao Zheng nodded at his words and noted it down in the document.

The Algorithm Group Leader closed the training panel and checked the time: 3:00 PM.

Chen Wen's side should have received the PCB prototypes by now.

Right then, his phone vibrated.

It was a message from Chen Wen.

"The boards are back, SMT just finished.

Testing the first revision tonight.

Why don't you come over tomorrow morning to align the interfaces?"

"No problem. I'll bring the test scripts over directly."

After replying, the Algorithm Group Leader leaned back against his chair.

A five percent confusion rate, eighteen percent for completely silent and motionless scenarios.

These numbers were bought with two iterations, over five hundred precision-labeled samples, over six thousand augmented samples, and an added LSTM temporal branch.

Although not ideal,

it was far better than before.

...