Robots need data, not just models: human video, fleet sharing, and a $90M seed bet

No free corpus

On Sequoia's Training Data podcast on October 20, Mecka CEO Josh Gao stated the physical-AI data problem plainly. As the podcast host put it, "unlike LLMs, where we had so much free human data from the internet, robotics is fundamentally data constrained" — and Gao breaks what data there is into three kinds: "Fundamentally, a lot of this robotics data is three things. It's pixels, it's actions, and it's logic." Mecka's answer is to run data collection as an industrial operation: the company creates, processes, and licenses data for robotics foundation models, defining egocentric data as video of humans demonstrating tasks from a first-person point of view. m:…6493

The operation is deliberately physical. Mecka employs a dual-approach strategy: a consumer mobile app for lightweight tasks — its earliest pilot had contributors recording themselves cooking and cleaning — and a B2B engine with custom hardware, including tactile gloves that capture force data during human task performance. The company says it has reached millions of hours of collected data and manages hundreds of petabytes, and Gao describes the industry-standard training mix as roughly a million hours of human egocentric data, tens of thousands of hours of UMI data, and a small amount of last-mile teleoperation and in-domain data. m:…6493

Fleets that teach themselves

The second route lets deployed robots generate their own data. In a paper published October 20, researchers introduce collective experience learning (CEL), "a learning paradigm in which deployed robots share their execution experience to improve a common policy" — without additional expert demonstrations. A fault-tolerant asynchronous training system coordinates experience collection, policy optimization, and distribution despite local interruptions in the fleet, while failure-guided retrieval pulls relevant successful trajectory segments from shared experience as supervision. m:…8754

The reported numbers: success rates of 93.3–99.2% across five real-world manipulation tasks, evaluated over 120 trials per task, and on multi-stage phone packaging, learning with 96 robots increased success from 79.2% to 98.3%. The authors also report that with collection time per robot held fixed, larger fleets generally achieve higher success rates — scale is the training strategy here, not the bottleneck. These are author-reported results, but the mechanism is the point: every deployed robot becomes a data collector, and the fleet's failures become the curriculum. m:…8754

Human video instead of robot labels

The third route sidesteps the most expensive input entirely. Action-labeled robot videos "directly supervise control but are costly and limited in diversity," so the AffordanceWAM paper — also published October 20 — uses affordance representations from human videos to jointly predict future scenes and robot actions. Human videos supervise the world streams while robot trajectories provide action supervision, enabling transfer without human action labels or motion retargeting. The scaling result is the one to watch: on RoboCasa, performance improves monotonically as affordance-annotated human video scales, and the model shows consistent gains over RGB-only and robot-data-only baselines across RoboCasa, CALVIN, and real-world manipulation. The abundant, unlabeled human side of the internet becomes usable robot fuel — no action labels required. m:…6124

The capital vote

Capital is voting with the data thesis. Dream Labs raised $90 million in seed financing, announced October 19 via Kindred Ventures — "Dream Labs' $90M seed financing round, led by B Capital, alongside Kindred Ventures" — to build what the company describes as the brain for everyday robots. The round drew participation from A.Capital, Unusual Ventures, SV Angel, Wing VC, and Scribble Ventures, plus individual founders and technologists. m:…2985

Read together, the funding and the research suggest where investors think the moat is forming: not in model architecture but in the data pipelines and collection infrastructure. Three sourcing strategies are converging — sensorized humans, fleet experience-sharing, and human video repurposed through affordances — and the VC read is that whichever team industrializes the data loop fastest wins the foundation-model race in robotics. The honest caveat comes from Gao himself: viability comes down to what commercial use case each scale of data unlocks, and general-purpose robots for complex, unconstrained environments are still some distance away. Data first, generality later. m:…6493