Robot Teleoperation and Demonstration Data Collection Statistics (2026): 50 Data Points on Cost, Throughput, and Collection Methods

Collecting one hour of robot demonstration data in 2026 costs anywhere from about $3 (a Chinese data-collection worker filming household chores, per Rest of World) to $48 (Tesla's posted rate for Optimus data-collection operators, per Fortune) in operator labor alone, before hardware and review. The largest teleoperated dataset, AgiBot World, packs 1,001,552 demonstration trajectories — 2,976.4 hours — collected by 100 robots in a single 4,000 m² facility, while NVIDIA's synthetic pipeline generated the equivalent of 6,500 human-demonstration hours in just 11 hours of compute. We aggregated 50 statistics from arXiv papers (ALOHA, UMI, DROID, AgiBot World, EgoDex), official research from Physical Intelligence, Figure, and NVIDIA, and named reporting from Rest of World and Xinhua on China's data-collection centers.

01What Demonstration Data Costs in 2026

The spread in operator cost is the story: the same hour of demonstration work prices at roughly $3 in a Chinese living room and up to $48 on Tesla's Palo Alto campus — a 16× labor arbitrage that explains why collection volume is concentrating in China. Per-session economics remain opaque across the industry; Xinhua's figure of over 1,000 yuan (~$140) for a single collection session is one of the few published unit costs, which is why teams increasingly benchmark against industrial-scale annotated data production instead of building in-house.

Note: the Tesla pay range comes from an August 2024 job listing — the most recent published US operator rate we could verify.

02Collection Methods Compared: Rigs, Throughput, Hardware Cost

Hardware cost per collection seat spans two orders of magnitude — a sub-$300 GELLO leader arm to a $32,000 Mobile ALOHA — and throughput gaps are just as wide. Handheld and kinesthetic interfaces beat joystick-style teleoperation on both axes: UMI is 3× faster than SpaceMouse teleoperation, and kinesthetic teaching doubles collection speed at near-100% demonstration success. Rig choice is now a budgeting decision as much as a research one, which is where vetted sourcing of capture rigs, data gloves, and stereo cameras earns its keep.

Note: ALOHA ($20k) and GELLO (<$300) figures are original 2023 published costs; component prices have drifted since.

03Flagship Teleoperation Datasets

Even the field's flagship open datasets are small next to what one frontier lab uses internally: π0's 10,000+ hours is more than three times all of AgiBot World, and DROID — 12 months of work by 50 collectors — totals just 350 hours. Scale is coming from centralized robot fleets, not distributed labs. For how these volumes feed model training more broadly, see our robotics training data statistics.

Note: AgiBot World episodes typically run ~30 seconds versus DROID's 5–20 seconds — per-trajectory length differs across datasets, so trajectory counts are not directly comparable.

04Egocentric Human Video and Wearables

Human-worn capture skips the robot entirely — and the volumes show why labs are betting on it. Apple's EgoDex recorded 829 hours across 338,000 episodes with nothing but Vision Pro headsets, and adding human embodiment data lifts robot task performance by 34–228% over robot-only training.

Note: Ego4D numbers date to 2021–2022 — the most recent available for that dataset.

05Sample Efficiency: How Many Demonstrations a Skill Needs

The strongest finding in the field is a power law: policy generalization scales with the number of environments and objects, not raw demonstration count — past roughly 50 demos per environment, extra demonstrations barely move success rates. That reframes collection from a volume problem to a diversity problem, and it is why 500 curated hours were enough to train Figure's Helix. The efficiency race among vision-language-action models is now largely a data-curation race.

06The Data-Collection Workforce and China's Teleoperation Farms

China is industrializing demonstration collection the way it industrialized manufacturing: more than 40 state-owned data collection centers announced by December 2025, purpose-built halls of teleoperated humanoids, and municipal output targets in millions of entries a year. The US model looks different — fewer, higher-paid operators inside companies like Tesla and 1X. Organizing collection across real factories, hotels, and kitchens rather than staged halls is the harder version of this problem — it's how we structure scene-partner collection at Cervo.

Note: for market context on the humanoid fleets these farms feed, see our humanoid robot market statistics.

07Sim-to-Real and Synthetic Data Generation

Synthetic generation is the only collection method whose unit economics improve with compute. NVIDIA compressed roughly 6,500 hours of human-equivalent demonstration into 11 hours of generation — and the mixed synthetic-plus-real recipe beat real-only training by 40%. Real demonstrations aren't going away; they're becoming the seed corn for much larger synthetic harvests.

Summary: Robot Demonstration Data by the Numbers

Methodology and Sources

This article aggregates 50 statistics from 27 primary sources: peer-reviewed and preprint robotics papers (ALOHA, UMI, DROID, AgiBot World, EgoDex, Data Scaling Laws), official company research and announcements (Physical Intelligence, Figure AI, NVIDIA, Apple), and named on-the-ground reporting on China's data-collection centers (Rest of World, Xinhua). We cite original studies rather than secondary blogs, exclude unsourced vendor cost estimates entirely, and flag figures older than two years. The Tesla pay range (2024 posting) and 1X NEO autonomy estimate are widely reported but rest on company statements rather than published studies.

• ALOHA — Learning Fine-Grained Bimanual Manipulation (Stanford, 2023)
• Mobile ALOHA (Stanford, arXiv 2401.02117, 2024)
• ALOHA 2 (Google DeepMind, arXiv 2405.02292, 2024)
• GELLO (arXiv 2309.13037, 2023)
• Universal Manipulation Interface — UMI (arXiv 2402.10329, 2024)
• DexCap (Stanford, arXiv 2403.07788, 2024)
• DROID (arXiv 2403.12945, 2024)
• Open X-Embodiment (arXiv 2310.08864, 2023)
• AgiBot World Colosseo (arXiv 2503.06669, 2025)
• RoboMIND (arXiv 2412.13877, 2024)
• EgoDex (Apple, arXiv 2505.11709, 2025)
• Ego4D (Meta AI, arXiv 2110.07058, 2022)
• EgoMimic (Georgia Tech, arXiv 2410.24221, 2024)
• HumanEgo (arXiv 2605.24934, 2026)
• Data Scaling Laws in Imitation Learning (ICLR 2025, arXiv 2410.18647)
• KineDex (arXiv 2505.01974, 2025)
• Open-TeleVision (arXiv 2407.01512, 2024)
• Figure AI — Helix (2025)
• Physical Intelligence — openpi / π0 (2024–2025)
• NVIDIA — Isaac GR00T N1 announcement (2025)
• Rest of World — How China is using human labor to win the humanoid robot data race (2026)
• Rest of World — In Chinese data factories, workers teach humanoid robots boring tasks (2026)
• Xinhua — Inside China's robot boot camp (2026)
• Fortune — Tesla hiring motion-capture workers at $48/hour (2024)
• eWeek — 1X NEO home robot (2026)
• The Robot Report — AgiBot releases humanoid manipulation dataset (2025)
• Apple Newsroom — Vision Pro availability and pricing (2024)

Last updated July 2026. We update this page quarterly as new data becomes available.