01 / Milky Way · Pre-training
Sovereign intelligence starts with the data.
We start from first principles: what must a model learn before a national team specializes it? Our data research focuses on building a useful, traceable foundation for physical intelligence.
Research agenda
Proposed work. Experimental results are forthcoming.
Begin with the physical problem
A photograph captures appearance. A sequence records change. An interaction links a decision to what happened next. Our research asks which combination gives a model reusable knowledge about the physical world.
Curate for the capability
Our proposed data curriculum brings together appropriately licensed video, robot trajectories, and simulation. We will vary their mixture against a defined capability target, track quality and duplication, and preserve held-out tasks for evaluation.
Provenance is part of the foundation
A national team needs to know what it can continue using. The intended record for each source includes origin, permitted uses, transformations, and known restrictions. A useful checkpoint should arrive with a clear account of the inputs and dependencies behind it.
Measure what each source contributes
Our planned comparisons hold the adaptation budget fixed, change one part of the training mixture, and measure transfer to unfamiliar settings. The result should explain both capability and resource cost, including cases where additional data did not help.
Our boundary
Orion focuses exclusively on pre-training. Evaluation and small adaptation probes help us assess the foundation; national teams lead subsequent post-training, specialization, integration, and deployment.
The foundation should be useful, understandable, and independently developable.