Tim Li spotted the problem while watching roboticists struggle to train their machines. "AI has read almost everything humans ever wrote," the DeepReach CEO noted in a mid-2026 LinkedIn post. "But it has barely watched how humans actually work."
His Y Combinator-backed startup reported having 475 devices in the field as of its September 2026 directory listing, each one strapped to mechanics, cooks, farmers, and factory workers whose movements have never been systematically recorded. In August, three months after launching its field network, DeepReach announced signing what it describes as a multi-million dollar data contract with a frontier AI lab. By September, the company reported generating nearly 150,000 video clips in the past three months and counted multiple robotics firms as customers, according to its Y Combinator directory listing.
The thesis is straightforward: Foundation models transformed language AI by training on vast text corpora scraped from the internet. Physical AI has no equivalent library. If robots are going to move beyond pre-programmed routines, someone needs to build one.
That gap showed up starkly in research from Capgemini's institute. Surveying organizations between January and February of this year, the consulting firm found 56 percent cited insufficient training datasets as physical AI's primary constraint. Another 43 percent flagged inadequate data infrastructure for training AI-powered robots. The six-person DeepReach team is wagering it can solve both by distributing cameras to local entrepreneurs rather than building a centralized data factory.
The model feels almost artisanal in an industry dominated by hyperscale labs. More than 150 entrepreneurs and over 1,000 local domain experts have signed on, the startup says, fanning out across Malaysia, the Philippines, Vietnam, Indonesia, and Thailand. Each partner recruits skilled workers, hands them stereo cameras, and captures egocentric video of tasks that have resisted automation: welding seams, folding dough, pruning crops, assembling circuit boards.
A Market Awakening
Industrial robotics posted its second-best year on record in 2024, with 542,000 units installed globally, according to the International Federation of Robotics. The United States added 38,000 units in 2025, an 11 percent jump over the prior year. China's operational stock crossed two million robots in 2024, the IFR reported.
Most of those machines still follow fixed scripts. They pick, place, weld, and paint in tightly controlled environments. Flexibility remains elusive. NVIDIA CEO Jensen Huang framed the challenge bluntly at the company's June 2024 Computex keynote, calling physical AI "the next wave" and describing it as "AI that understands the laws of physics, AI that can work among us."
NVIDIA shipped its GR00T N1 humanoid robot foundation model in March 2025, then released an updated version alongside a synthetic motion data blueprint two months later. The message to the robotics industry was clear: the tools exist. The bottleneck is data.
Early efforts to assemble training corpuses have been fragmented and relatively small. Open X-Embodiment, a consortium of 21 institutions announced in October 2023, pooled 60 datasets covering 527 skills across 160,266 tasks. DROID, published at the Robotics: Science and Systems conference in 2024 and updated through early 2025, contributed 76,000 teleoperated trajectories spanning roughly 350 hours across 564 scenes. Stanford's AI Index report this past April documented progress on robot foundation models but underscored data scarcity as a persistent obstacle.
Figure, the humanoid robotics startup, disclosed in September that its Index dataset generates "35 minutes of data every second" and announced a partnership with Nscale for access to up to 100,000 GPUs on NVIDIA's Vera Rubin infrastructure. A year earlier, Figure partnered with Brookfield to capture human video across the real-estate giant's global portfolio, aiming to build what the companies called "the world's largest and most diverse humanoid pretraining dataset."
Universal Robots and Scale AI unveiled a force-aware imitation learning system at NVIDIA's GTC conference in March, integrating Scale's Physical AI Data Engine with Universal's e-series cobots, which the Danish manufacturer has deployed in over 100,000 locations worldwide. The system synchronizes robot motion, vision, and force feedback data. Anders Beck, Universal's VP of AI Robotics Products, said it gives developers "direct influence over how the robot physically interacts with the world" by incorporating torque control and force sensing.
Scale AI now claims to generate over 1,000 hours of robotics data per day, a company general manager told The Information. Google DeepMind's AutoRT project ran 52 robots across office settings in early 2024, gathering 77,000 trials covering 6,650 tasks, the lab announced that January. The scale has grown rapidly, but centralized data collection in controlled environments has limits. Real-world variability is harder to capture.
The Distributed Model
DeepReach describes itself as a "platform where entrepreneurs build local data businesses serving Physical AI." It supplies the cameras, software, quality assurance infrastructure, and payment rails, according to its LinkedIn profile. Entrepreneurs handle recruitment and logistics on the ground.
The approach echoes gig-economy platforms but targets skilled labor rather than commodity tasks. A welder in Penang or a machinist in Ho Chi Minh City becomes a data contributor, paid per clip that meets quality thresholds. The company has posted references to filtering standards across datasets it calls EgoView, EgoVerse, and EgoDex. Its Hugging Face organization page lists "Egocentric Manipulation" and "DROS — Distributed Robot Operating System" as research areas, though neither has been formally published.
Cost dynamics appear to favor the model. Analysis from the Silicon Valley Robotics Center claimed this past June that high-quality teleoperated demonstration costs fell from roughly $340 per hour in early 2024 to $136 per hour in the fourth quarter of 2025. The methodology was not independently verified, and DeepReach has not disclosed its own per-hour economics. Accessing labor markets with lower wage structures and eliminating centralized facility overhead could push costs lower still, though quality control becomes more complex.
Other startups are testing variations on the theme. Sensei, also a Y Combinator company, positions itself as "Scale AI for robotics data" with a paid operator network. Field Motion and Robgence offer egocentric video capture and teleoperation infrastructure with multi-modal annotations. Noitom Robotics released HiPHI, billed as one of the largest high-precision human motion datasets, at the World Robot Conference in August, targeting humanoid learning applications.
The competition suggests validation of the core insight: generalist robot models require far more diverse training data than any single lab or company can generate internally. Whether distributed networks can maintain the data quality and consistency needed for effective model training remains an open question.
Regulatory Headwinds
Capturing workplace video at scale introduces legal complexity. The EU's AI Act transparency rules took effect in August, with high-risk system obligations phasing in starting December 2027. Industrial robots in safety-critical settings may fall under high-risk classifications, requiring dataset documentation, risk assessments, and post-market monitoring.
California's Consumer Privacy Protection Agency adopted updated CCPA regulations effective this past January, mandating risk assessments for high-risk processing and automated decision-making opt-outs. Illinois biometric privacy litigation continues to generate settlements; a recent HireVue case covered participants through late June. Egocentric workplace video can implicate employee data rules and biometric thresholds, particularly when facial or gait recognition is involved.
ISO revised its collaborative robot standards in 2025, superseding 2011 editions and integrating force-threshold guidance from an earlier technical specification. The United States harmonized via ANSI standards reaffirmed in 2026. Physical data capture near human-robot collaboration zones must observe updated safety modes and force limits, adding operational constraints.
DeepReach has not publicly detailed its consent protocols or data anonymization practices. For a startup moving fast across multiple jurisdictions, compliance infrastructure typically lags deployment. That gap represents risk, though perhaps not the existential kind in a market where regulatory frameworks are still taking shape.
The Path Forward
The September Y Combinator directory listing said DeepReach aims to scale "to 10,000+ devices across 50 countries." The company declined to disclose funding details beyond its YC participation. A team of six running a distributed network across multiple continents suggests lean operations and tight margins, or perhaps both.
Market projections for humanoid robotics vary wildly. Goldman Sachs forecast growth from $6 billion to $38 billion by 2035 in blue-sky scenarios published in early 2024. Omdia projected global humanoid shipments exceeding 10,000 units by 2027 and reaching 38,000 in 2030. UBS estimated million-unit sales by 2030 in a November 2025 outlook. The wide variance reflects uncertainty about cost curves, dataset availability, and the difficulty of transferring skills from simulation to physical hardware.
Infrastructure is converging around open standards. Hugging Face released LeRobot v0.6.0 in July, adding world-model policies, unified evaluation tooling, and expanded vision-language-action model support. The framework integrates NVIDIA's GR00T family and third-party models, signaling a shift toward open benchmarks that could accelerate experimentation.
Toyota Research Institute CEO Gill Pratt said in a September 2023 press release that the institute's robotics work aimed at "amplifying people rather than replacing them." DeepReach's network of local experts suggests a parallel logic, though with a different motivation. The company is not building robots. It is building the library that will teach them.
Whether distributed data capture becomes the industry standard or a stopgap until synthetic data generation improves remains unclear. McKinsey projected in mid-2026 research that physical AI and robotics could create a trillion dollars in value by 2040, contingent on solving data readiness and simulation-to-real transfer. That timeline gives DeepReach and its competitors years to refine their models, or to be outpaced by labs that crack synthetic generation at scale.
For now, the wearable cameras keep recording. Mechanics weld, cooks chop, farmers prune. Somewhere in those hundreds of thousands of clips is the training data foundation models need to move beyond narrow tasks. The question is whether human demonstration at scale provides signal that simulation cannot yet replicate, or whether capturing the world as it is will always lag behind imagining the world as it could be.
