Robot Scaling Claim Caution / 幂律现象与 scaling law 谨慎

Definition

Robot scaling power-law caution is the practice of reporting that robot capability improves with data, compute, and data diversity without claiming that a validated scaling law has been established.

Current Synthesis

The bounded source says the team sees performance gains from data diversity, compute, and data volume, and would call what it observes a power-law phenomenon rather than a scaling law, because a real scaling law requires serious and sufficient experiments. That caution is paired with a data strategy: self-collected UMI data is cost-controlled, high-quality, and reported as more than 90% usable, which the company says lets it run controlled experiments across model-parameter and data-ratio settings. The same source is skeptical of million-hour egocentric data on cost grounds, since the raw footage may be cheap while annotation and training are expensive and the payoff is unverified, and it treats zero-shot task performance as a signal of pretraining quality rather than proof of general capability.

Key Claims

• Power-law-like improvement is not the same as a validated scaling law, and the source deliberately keeps the weaker claim.
• Data diversity, compute, and data volume are all reported as capability levers.
• Controlled scaling experiments depend on owning a data pipeline whose cost and usability are known.
• Large-scale human first-person data can be cheap as raw material yet expensive once annotation and training are included.
• Zero-shot task success is used as a pretraining-quality signal, not as a finished capability claim.
• The source separates scaling evidence from marketing, and ties its caution to the absence of public benchmarks and reproducible evaluation.

Evidence

• Scaling-caution evidence: 于是转身向具身走去｜对话王家伟：24 岁的具身智能首席科学家 says the team sees power-law phenomena but avoids claiming a scaling law because verification would require serious, sufficient experiments.
• Controlled-experiment evidence: 于是转身向具身走去｜对话王家伟：24 岁的具身智能首席科学家 says self-collected UMI data is cost-controlled and over 90% usable, enabling experiments across different parameter counts and data ratios.
• Ego-data cost evidence: 于是转身向具身走去｜对话王家伟：24 岁的具身智能首席科学家 argues that million-hour egocentric data may have low raw-material cost but high annotation and training cost with unverified effect.
• Zero-shot evidence: 于是转身向具身走去｜对话王家伟：24 岁的具身智能首席科学家 reports zero-shot task performance before release and connects it to data quality rather than to a general-capability conclusion.

Counterevidence & Qualifications

The page records one company’s framing; no scaling curves, experiment counts, or model sizes are disclosed, and the power-law observation is unverified outside the company. The claimed 90%-plus usability is also self-reported. The caution should therefore be read as a discipline about claiming too much, not as evidence about the true shape of robot-learning returns.

What Changed

• Created the concept from the interview’s power-law-versus-scaling-law distinction.
• Added the data-ownership condition that the source says makes controlled scaling experiments possible.
• Recorded the cost objection to million-hour egocentric data as part of the scaling decision.

Related Concepts

• Robot Data Scale Up - broader data-scaling bottleneck this concept qualifies.
• Robot Generalization Performance Tradeoff - capability tension that scaling experiments are meant to measure.
• Egocentric Robot Data - large-scale first-person route whose cost-effectiveness the source questions.
• Embodied Native Foundation Models - robot-native model route that still depends on usable data scale.
• Frontier Model Scaling - language-model scaling discourse whose confidence this concept refuses to import wholesale.
• Robot Evaluation Problem - missing measurable evaluation behind the refusal to claim a law.