September 25, 2026 · @keval

2 · The numbers@1x.png

Moving quality checks from the data center to the collection site can lift usable yield from 68% to 85% and save roughly $8,000 a month in wasted collection cost. All figures in this case study are illustrative assumptions for a mid-sized wearable-kit program unless marked otherwise.

Summary
Problem About 1 in 3 recorded hours is unusable, and failures surface ~18 days after recording
North star Usable yield = usable hours ÷ recorded hours
Key insight The cost of a failure depends on how late it is detected, not just on how often it happens
Solution (MVP) Automated pre-session check + on-site exit validation + offload integrity gate
Target Usable yield 68% → 85% in one quarter, with setup time up by no more than 3 minutes
Role shown Problem framing, metric design, root cause analysis, prioritization, experiment design

1. Context: why real-world data matters

Physical AI can only be as good as the real-world data it learns from, and that data is expensive to collect. BCG defines physical AI as AI that perceives, reasons and acts in the real world through robots, machines and vehicles. It treats embodied AI as the narrower research idea of an agent with a body.

The market is moving fast. Cisco cites growth from $1.5 billion in 2024 to over $15 billion by 2032. It lists data collection and training as the first stage of the sense-plan-act loop.

Simulation alone is not enough. Cisco names the sim-to-real gap as a core challenge: models trained in simulation struggle with real friction, lighting and sensor noise. That is why companies still pay humans to record real tasks.

Quality is the bottleneck, not just volume. EY tells leaders to treat data as the fuel for autonomy and stresses that it must be structured, synchronized and high quality.

Public datasets show how this data is collected at scale:

Collection method Well-known example Typical failure
Egocentric video Ego4D, Ego-Exo4D Hands out of frame, motion blur
Wearable multimodal kit This case study's setting Camera–IMU sync drift, dead sensors
Teleoperation ALOHA, DROID Failed or jerky demonstrations
Handheld gripper UMI Camera tracking loss, lens calibration
Pooled robot data Open X-Embodiment Inconsistent formats, missing metadata

2. The pipeline and the problem

A recording passes through five hands before anyone checks whether it is usable, so failures are found weeks too late to fix.

flowchart LR
A[Collector records<br/>wearable kit] --> B[Drive staging<br/>at site]
B --> C[Ship drives to<br/>data center]
C --> D[Upload to S3]
D --> E[QA review]
E -->|~15-20 days later| F[Usable or discarded]

The setting: a program of 40 collectors records household and workplace tasks with a head camera, wrist cameras and IMUs. It records about 1,200 hours a month.

Problem statement: 32% of recorded hours are discarded at QA. By then the collector has left the site, the task setup is gone and the only fix is a full re-record.

Cost of one failed hour

Cost item Per recorded hour (USD, illustrative)
Collector pay and travel $18
Kit wear and amortization $6
Drives and shipping $5
Upload and storage $3
QA review time $8
Fully loaded cost $40 
\text{Monthly waste} = 1{,}200 \times 0.32 \times 40 \approx 15{,}360 \text{ USD}

That excludes the harder cost: re-recording delays dataset delivery to the ML team or customer.