A look at data labeling startups like Objectways, whose workers record and annotate repetitive tasks like folding towels to train AI robots for physical tasks
where young engineers strap GoPros to their foreheads and fold laundry or pack boxes to teach humanoid robots how to do chores. https://www.latimes.com/... [video] Nilesh Christopher / @nilchristopher : It has also driving some ingenious ways of real-world data collection @aliniikk 's Micro1 pays people upto $50 an hour to wear Meta Ray ban smart glasses to capture everyday actions. Figure AI does something more crazy👇 https://x.com/... Nilesh Christopher / @nilchristopher : “Sometimes we have to delete nearly 150 or 200 videos because of silly errors in how we're folding or placing items,” said Kumar, an engineering graduate who has worked at Objectways for six years. His firm sent 200 towel-folding videos to its client in the United States [video] LinkedIn: Ravishankar Rajalingam : Excited to share that Objectways was featured in the Los Angeles Times for our pioneering work in training AI and humanoid robots with real-world human data! …
Context & Ripple Effects
AI training-data work has long relied on distributed human labelers, from Google’s quality-rater contractors to the broader tasker workforce documented around Scale AI. This coverage shifts that labor from classifying digital content to capturing the physical steps a robot must imitate.
The move also comes as Scale AI sought higher-margin AI tools beyond its large contractor base in its push beyond labeling services. For physical AI, the collection process itself—cameras, demonstrations, annotation, and review—becomes part of the product.
First-order effects
- Objectways and similar vendors can sell curated recordings of routine physical actions as training inputs for humanoid-robot developers, while workers are paid to perform and document those actions.
- High rejection rates for incorrectly performed recordings make quality control an immediate bottleneck: each usable task sequence requires more worker time and review than a simple completed video.
Second-order effects
- Robot developers will have to weigh collecting demonstrations in-house against buying specialized datasets; vendors that can deliver consistent annotations and task execution gain a clearer advantage.
- Wearable-camera approaches, including Micro1’s use of Meta Ray-Ban glasses, broaden the pool of potential data contributors but also make capture standards and review processes more consequential.
Third-order effects
- If this model scales, training data for physical AI may become a differentiated supply chain rather than a generic labeling service, with repeatability and task coverage determining dataset value.
- The pattern points toward more workflow-native data businesses: human work is increasingly captured as reusable training material, though the economics will depend on whether data quality improves faster than collection and review costs.
The trend: Physical AI is creating demand for specialized, quality-controlled demonstrations of real-world work, extending the data-labeling industry from digital judgment to embodied tasks.