How PsiBot Uses 100,000 Hours of Human Data to Power Embodied Intelligence

In 2026 the term “world model” has become a buzzword in the embodied‑intelligence field, as companies claim that learnable environment models can boost robot training efficiency. PsiBot’s co‑founder Chen Yuanpei stresses that world models are merely tools for data migration, not the core focus; the real question is whether large‑scale human operation data can be transformed into robot training data.

Before PsiBot, Chen had already explored using hand‑movement data for dexterous manipulation, publishing the work at CoRL 2024. He now asserts that at the 100,000‑hour scale, human data can largely replace data collected from real robots.

The company pursues three parallel data streams: (1) exoskeleton‑glove data that mechanically captures hand and arm motions without relying on IMUs, offering high precision and full bilateral freedom; (2) pure visual ego‑data recorded by head‑ and wrist‑mounted cameras, which is cheaper and more scalable but less precise; (3) a combination of both to maximize coverage of real‑world labor scenarios such as logistics, warehousing, checkout, and factory work.

Compared with the traditional teleoperation route—where operators control robots or shadow arms in a dedicated “material‑field”—the glove‑based human‑centric approach sacrifices immediate transfer efficiency for dramatically higher data‑scale potential. Teleoperation suffers from high collection cost, heavy equipment, venue dependence, and the need for trained operators, limiting its scalability.

PsiBot’s solution is a two‑module system: W0, an action‑conditioned world model that predicts the next state given the current state and action, and R2, the policy that ultimately runs on the robot. During training, W0 acts as a learnable simulator; R2 iteratively explores within this simulated environment via reinforcement learning, converting human dynamics into robot dynamics and generating new training data that feeds back into R2. In deployment, W0 is removed and only R2 runs on the robot.

Data quality is filtered automatically: a data point is retained only if the world model can successfully convert it and the resulting policy can execute without failure. As model capabilities improve, the filtering boundary shifts, allowing more diverse human data to be used.

The company released the SynData dataset on Hugging Face (≈1.46 k downloads as of May 13 2026), a multimodal collection covering vision, language, and action, captured with the exoskeleton glove and complemented by raw hand data. SynData serves as the foundation for training the W0‑R2 pipeline and is openly available for research on action modeling, manipulation learning, and multimodal intelligence.

Strategically, PsiBot distinguishes three development stages: (1) a capacity stage where revenue comes mainly from hardware (gloves, capture systems, material‑field construction); (2) a policy‑tuning stage where robot policies are adapted to specific tasks and environments; and (3) a future base‑model stage, which the company does not yet consider imminent. The “small‑stack” approach means PsiBot builds core components (data pipeline, world model, policy) in‑house while sourcing peripheral hardware such as tactile sensors and gearboxes.

Chen argues that while real‑robot teleoperation data remains valuable for calibration and fine‑tuning, it is not the sole fuel for robot learning. If human data can be collected at scale and the migration pipeline is effective, it can substantially replace teleoperation data. However, human data must pass through the full pipeline—capture, world‑model conversion, reinforcement learning, filtering, and policy training—to become usable.

Regarding industry trends, Chen is skeptical about simulation breakthroughs solving the physical‑world gap and believes that the most likely path to disproving the human‑data route would be a massive simulation advance or a company with enough resources to run a true‑robot data flywheel. He concludes that the real moat lies in the data pipeline and organizational capability, not in secret algorithms.