A world model that learns the body,
action-conditioned, self-supervised
Every wearer-hour scores itself —
prediction error is the only reward
Upstream of every humanoid,
the substrate for physical AI
Most models need humans to label data or engineer rewards. PROPRIO-1doesn’t. It predicts the latent consequences of motion — the next state of the body in contact with the world — and the gap between prediction and reality is its own training signal. Every wearer-hour is automatically scored, with no annotation pipeline.
Self-supervised world models work — but they’ve been starved of the one modality that matters most for physical intelligence: whole-body, contact-rich, force-labeled human movement. Vision shows what an action looks like; force reveals why it worked. PROPRIO-1 is built on paired vision-and-force capture — the same body, the same instant, both streams time-aligned — so it can learn the mapping between appearance and the invisible forces that drive successful action. That bridge is what pure-vision models structurally cannot learn.
Video pre-training
Unlabeled robot video to adapt
Task-specific reward (zero-shot control)
Meta FAIR, June 2025. PROPRIO-1 applies this self-supervised approach to data it has never had: whole-body human contact and force.
PROPRIO-1and HALO’s data improve each other on one loop. The model learns from captured movement; its prediction error tells us which data is most valuable; that data compounds the moat.
Roadmap — the model is the long-term moat; data licensing is the near-term engine. See Data →
The first milestone is demonstrating the prediction-error loop on real captured data — not benchmarking against frontier models. We’ll publish results when the loop is up.
Prove the prediction-error training signal works on PROPRIO-1's own captured data — whole-body, force-labeled, paired with vision.
Expand wearer-hours, deepen vision + force time-alignment, and grow the curated training corpus.
Expose the latent world model as a control and evaluation substrate for humanoid and embodied-agent teams.