The Hidden Labour Behind AI: Inside India’s $3/hour Workforce Training Humanoid Robots-and What It Means for the UK

Indian workers paid $3/hour to train humanoid robots: what’s actually happening

A viral Reddit post claims Indian workers are being paid a little over $3 per hour to wear head cameras and limb sensors while performing everyday tasks. The footage – cutting mangoes, washing dishes, moving through homes and textile factories – is used to teach humanoid robots human-like movements.

The post names Objectways as the data collection firm, with a subcontractor (Qanat Consulting Services) reportedly managing 2,000 workers fitted with tracking bands. One worker records up to 90 clips a day, at around four minutes each, filmed in furnished studio apartments and factories. Objectways’ clients are said to include Fortune 500 companies. A quoted line from CEO Ravi Shankar frames the aim as automating some jobs so people can focus on “more useful tasks”.

There’s also a headline-grabbing forecast: Morgan Stanley predicts the global humanoid robot population could exceed one billion by 2050. That’s a big number – worth treating cautiously – but it mirrors a wider shift: data-intensive robotics is moving from lab demos to industrial-scale data collection.

Source: the original Reddit thread.

Key details from the post

Metric	Value
Hourly pay	Just over $3/hour
Clips per worker	Up to 90/day
Clip length	~4 minutes
Workers managed	2,000 (via subcontractor)
Collection sites	Textile factories, furnished studio apartments
End clients	Fortune 500 (not disclosed)
Robot forecast	1B humanoids by 2050 (Morgan Stanley, per Reddit)

How head cameras and motion sensors train humanoid robots

Humanoid robots need “embodied” training data: sequences of body and hand motions, associated with tasks and environments. Head-mounted cameras provide a first-person view of objects, surfaces, tools, and spatial context. Motion sensors on limbs track joint angles and trajectories. Combined, this supports imitation learning – models learn to reproduce sequences of actions – and helps build richer datasets for planning and control.

Teams then run data annotation: labelling segments (start/end of “pick up cup”), tagging objects, and linking sensor streams. “Annotation” is simply structured labelling that makes raw footage useful for training. The post suggests this work is being industrialised in India – with scale, standardised environments, and performance targets.

Ethical trade-offs: low-cost human labour training robot labour

The economics are stark: low wages by Western standards, physically repetitive work, and intimate recordings of homes and workplaces. At the same time, participants are being paid above typical data annotation rates in some markets – but the post does not disclose comparisons, conditions, or benefits. Without that context, it’s hard to make a fair assessment.

“Certain jobs must be automated so humans can focus on more useful tasks.”

There’s a valid argument that offloading drudgery is a social good. The problem is sequencing: people asked to do repetitive data work at low pay now may be the first displaced later, unless the industry offers pathways into higher-skilled roles (tooling, QA, supervision). That requires intentional design and transparent standards.

Why this matters for the UK: compliance, procurement, and competitiveness

UK companies building or buying robotics systems are part of this supply chain. If training data includes identifiable people, domestic interiors, or workplace footage, UK data protection rules bite – even if the capture happened abroad. That means documenting provenance, consent, and transfer mechanisms, and ensuring data minimisation and security.

Before procuring models or datasets trained on human-capture streams, UK buyers should run a Data Protection Impact Assessment (DPIA) and demand clear data lineage. See the ICO’s guidance on DPIAs under UK GDPR.

There’s also an ESG thread. Investors and customers increasingly expect transparency on labour practices within AI supply chains. Low-cost annotation work is not inherently exploitative – but opaque sourcing undermines trust. If a model can only be competitive by externalising labour and privacy risks to the most vulnerable, that’s a red flag.

UK buyer checklist for robot training data and suppliers

• Data provenance: where, when, and by whom was footage collected? Environments staged or natural?
• Consent: were participants informed about downstream commercial use? Are faces, voices, and personal items redacted?
• Worker safeguards: pay structures, breaks, ergonomic risks, and reporting channels. Independent audits?
• Privacy controls: on-device redaction, secure storage, retention limits, and deletion on request.
• Cross-border transfers: lawful transfer mechanisms and vendor subprocessor lists.
• Licensing: explicit dataset licence terms and redistribution restrictions.
• Model cards: disclosure of training sources and known limitations for downstream risk management.

Where humanoid robots could land first in the UK

Even if the “billion by 2050” forecast is aggressive, near-term deployments are likely in warehouses and logistics, light manufacturing, facilities management, and possibly in social care as assistive devices. The common thread is repeatable tasks in semi-structured environments, where safety cases and ROI are clearer.

Costs, availability, and performance for UK buyers are not disclosed in the Reddit post. Expect a long tail of integration costs – safety validation, insurance, workflow redesign – beyond the robot sticker price. Also remember the hidden infrastructure footprint (compute, training, inference). For a related angle, see my explainer on AI’s water and energy use in data centres.

Interpreting the “1 billion humanoids by 2050” prediction

Hardware forecasts decades out are notoriously fragile. Constraints range from battery energy density and dexterous manipulation to safety regulation and public acceptance. Even if unit costs fall, capability and reliability are the limiting factors. Treat the number as a provocation, not a certainty.

“Morgan Stanley predicts the global population of humanoid robots will exceed 1 billion by 2050.”

What is clear is the data pipeline: massive human-captured motion datasets are becoming an industry standard input. The question for the UK is whether we participate with strong safeguards – or import opaque models and inherit their risks.

Bottom line: practical takeaways for UK developers and product teams

• Data quality beats data quantity: prefer well-instrumented, consented capture over noisy, privacy-sensitive footage.
• Interrogate suppliers: ask for data lineage, redaction processes, and worker welfare evidence. “Not disclosed” is a sign to pause.
• Run DPIAs early: if your product benefits from human-capture data, build privacy by design from the outset.
• Plan for safety and integration: field performance will hinge on edge cases, not glossy demos. Budget for testing and iteration.
• Stay balanced: automation can remove drudgery and lift productivity, but only if we align incentives across workers, buyers, and users.

If you want to examine the original discussion and claims in context, here’s the Reddit post. The specifics above are drawn directly from that thread; many figures and client names beyond those mentioned are not disclosed.