Robot Training Data Crisis: From Low Wages to Collapsing Data Credibility

Photo by zhang kaiyv / Pexels

---

I noticed an interesting detail: Daily wage of 250 RMB, 5 cameras, 6 hours of effective footage. This data collection efficiency, if placed in the risk control domain, is a typical "low signal-to-noise ratio" scenario—a tiny proportion of valid signals amidst massive raw data, with truly valuable information often drowned by noise. More critically, this data ultimately feeds into robot models, and models are far more sensitive to data quality than humans are to household chores.

1. The "Gold Content" of Data Collection

First, a basic comparison:

Dimension	Human Perception	Robot Model Requirement
Action Consistency	Doesn't matter, different poses daily	Strict consistency required, otherwise model gets confused
Environmental Lighting	Highly adaptable	Extremely sensitive to lighting changes
Background Interference	Automatically filtered	Prone to learning background noise as features
Anomaly Handling	Instinctive reaction	Requires extensive negative sample coverage

Zhang Yue collects 6 hours of effective footage a day, but the "clean" data usable by models might be less than 30 minutes. Why? When humans do housework, they subconsciously adjust body posture, block cameras, and create noise via clothing friction. These actions seem "normal" to humans but are catastrophic feature shifts for robot models.

In risk control, I've seen many similar cases: Collected samples seem to cover all scenarios, but once the model goes live, it encounters insufficient rule coverage because the distribution of "positive samples" (normal behavior) and "negative samples" (anomalous behavior) in training data deviates severely from the real environment. Household data collection is the same—Zhang Yue's movements at home may differ from people of different heights, furniture layouts, and clothing materials in real households by more than the model's tolerance threshold.

2. Engineering Challenges of Data Annotation

Companies like Lightwheel Intelligence earn margins from "data cleaning + annotation." But the issue is that the complexity of annotating household data is far higher than autonomous driving.

• Autonomous driving annotates static objects (lane lines, pedestrians, vehicles) with clear boundaries.
• Household actions annotate continuous motions (tying shoelaces, folding quilts) with blurry boundaries.

Take "folding a quilt" as an example; one action sequence includes:

• Grabbing corners → Shaking to unfold → Folding in half → Flattening → Folding again
• Keyframe identification requires annotating at least 5 time points for action start, end, and object state changes.
• Different quilt materials (cotton, down, blanket) cause variations in motion amplitude that need independent annotation.

Current industry practice involves annotators labeling frame-by-frame, but a 5-second action clip takes an average of 15 minutes for precise annotation. This means Zhang Yue's 6 hours of footage might require 72 hours of annotation time, costing far more than the 250 RMB/day wage.

I suspect Lightwheel Intelligence's profit model suffers from data quality inflation—to compress costs, they likely adopt a mixed strategy of "rough annotation + manual spot checks." This directly introduces massive annotation errors into training data, resulting in robots "learning" an action but missing key steps or being off by a few centimeters during execution.

3. Risks in the Data Supply Chain: A Black Market Perspective

As someone dealing with black markets daily, my first reaction to this industry chain is: The trust model for outsourced data collection is too fragile.

• Zhang Yue signs contracts directly, but large-scale data collection involves multi-layer subcontracting, ending up in small workshops.
• Small workshops might use AI-generated videos instead of real collection; since camera angles are fixed and actions are templated, it's hard to tell.
• The annotation phase also faces "volume padding": Annotators might use automated scripts for quick labeling, just to pass spot checks.

I've seen similar tactics in black market chains: Mixing 10% forged data with real data causes model performance to drop by less than 5%, but cuts costs by 90%. For robot household data, if end-users (robot companies) cannot verify the authenticity of data sources, suppliers have strong incentives to inflate quality.

Verification methods exist but are costly: Checking metadata for every video frame (timestamp continuity, natural lighting changes, biomechanical plausibility of actions) is exactly like analyzing behavioral trajectories to detect identity fraud in risk control. But the industry currently lacks even basic "data fingerprinting" mechanisms.

4. Viewpoint: The "Dam Lake" Effect of Data Quality

The current situation resembles internet ad traffic fraud in 2015: Advertisers poured money, media platforms padded volumes, and third-party monitoring was nominal. The difference is that ad fraud loses money, while robot data fraud loses industry trust.

The real bottleneck for robot training data isn't collection volume, but credible verification of data quality. When massive amounts of low-quality data flood the market, it leads to:

• Rising model training costs (need more data to offset noise).
• Lower ceiling for model effectiveness (dataset quality determines model upper limit).
• Data suppliers falling into "bad money drives out good" (companies insisting on high quality get eliminated by price wars).

Lightwheel Intelligence makes money now because the market is in a "data hunger phase" where any data has value. But once robot companies start conducting post-cleaning model performance reviews, they'll find data quality severely drags down performance. At that point, like in risk control, they will demand Data Quality Reports from suppliers (including metrics on annotation consistency, scenario coverage, anomaly ratios) rather than simply paying by the hour.

One-sentence summary of core view

Before robots learn to do housework, we must bridge the trust gap between "human inflation" and "machine screening" in data collection, otherwise selling...

Original Link: https://www.tmtpost.com/8062934.html