DeepReach, a six-person startup founded in 2025 and part of Y Combinator's Summer 2026 batch, has launched a business that sells recordings of real-world human work to companies training robots, according to its Y Combinator launch post. The company is arriving in a market where startups and robot makers are chasing the same scarce resource.
Figure runs its own in-house capture effort. XDOF and Mecka have both drawn investor attention.
What DeepReach says it built
Tim Li is the CEO and Chris Liu the CTO, according to Y Combinator's company directory. The Grangel directory lists Brad Flora as the YC primary partner.
The directory describes wearable stereo capture devices, designed by the company and used in warehouses, workshops, farms, kitchens and repair shops. It also says multiple frontier-model and robotics companies use the data in production. It does not name any of them.
The launch post is titled "DeepReach: Launch a real-world data business today." The page itself says only "3 months ago." A third-party index dates it to August 13, 2026, though the primary page does not confirm that.
Verification is thin. The YC directory and the launch page disagree on clip totals, partner counts and target countries, and neither carries a revision date. DeepReach's own website returned access errors during research, and no independently verified customer case study turned up.
One YC job listing, for a head of global data partner network operations, describes expansion across the United States, Latin America, Europe and Asia. Those are hiring ambitions. They are not confirmed operating countries.
Funding remains unclear
The materials reviewed did not disclose the size of DeepReach's round. Robinhood announced on October 1, 2026 that it had invested about $15.75 million across 46 YC S26 companies, in checks ranging from $100,000 to $600,000. That figure is an aggregate and should not be read as DeepReach's round.
Under YC's standard deal, per its deal page, a startup receives $500,000: $125,000 for 7% of the company, plus $375,000 on an uncapped MFN SAFE (a convertible instrument that converts to equity at a later financing). Team Ignite says in its own blog post that it is a DeepReach investor, so it cannot be treated as an independent evaluator.
Funding databases also call for caution. Dealroom's Deepreach.ai record pairs the robotics company's domain with the financing history of an unrelated French adtech firm at deepreach.com.
The competition
XDOF is the most prominent comparison. TechCrunch reported on September 4, 2026 that the company, only three months out of stealth, was in talks for a Series B at a valuation of roughly $1.2 billion. That was a reported negotiation, not a closed deal.
Mecka has announced a Series B, according to its company site. Its run-rate figure is not audited annual revenue, and its year-end number is a forecast.
Encord announced a $60 million Series C on February 26, 2026, led by Wellington Management, which brought its disclosed funding to $110 million. Its named customers include Woven by Toyota, Skydio and Synthesia. The company reported that its physical-AI revenue grew tenfold over the preceding 12 months, a figure now more than six months old. On October 20, 2026, it launched Data Labs and EMG-1, a dataset pairing 307 hours of field wrist-muscle signals with egocentric (first-person) video.
Smaller YC S26 peers are dividing up the problem. Hebbian Robotics sells APIs for data search, duplicate detection and monitoring of quality drift. 6thSense builds wearable rigs that capture pressure, grip, force, depth and hand pose, and argues that video alone leaves out tactile information.
Then there is the build-versus-buy threat from Figure's Index. Figure's own account is not an independent comparison of vendors, but it shows a robot maker running internal collection.
Questions about differentiation
Distributed, robot-free demonstration capture predates DeepReach, according to an arXiv paper (2402.10329). Berkeley's DROID work likewise treated distributed collection and environmental diversity as research priorities. Any edge for DeepReach would therefore have to come from access, collection economics, hardware, data quality or customer outcomes.
The evidence that geographic diversity helps is partial. One arXiv analysis (2604.07607) relied mainly on an offline action-prediction error metric rather than downstream robot task success. Another (2608.02580) found that human capture feeds a substantial reconstruction and alignment pipeline and is not necessarily usable as action supervision on its own.
What to watch
Physical Intelligence's October 2026 API workflow uses deployment failures to set the next collection priorities, which points toward targeted data over raw volume. Stanford's January 2026 robotics report argues that simulation stays valuable but needs grounding in the real world. ANYbotics' October synthetic-data case with NVIDIA, documented by Encord, shows synthetic data complementing real collection, not replacing it.
Independent capture networks face well-funded rivals and robot developers building private pipelines, Mecka's announcement indicates. For DeepReach, the next test is to name a customer and show a trained model that improved.
