Technical Whitepaper · RL²

Real-World Evaluation for Physical AI

In software, evals became the bottleneck — the sharpest AI teams are racing to build them. In the physical world, that layer doesn't exist yet. This paper lays out why real-world evaluation, not a bigger model, is the moat in Physical AI — and the loop we built to close it.

• Why robot models saturate benchmarks but still fail in the workcell
• The RL² loop: deploy → capture failures → evaluate against acceptance criteria → collect targeted data → redeploy
• "Acceptance criteria are the new PRD for robots" — how deployment gates compound into a durable asset
• The SVRC one-stack: hardware → data → evaluation, and where value actually accrues
• An honest map of what's automated today vs. the multi-year build ahead

Prefer the narrative version first? Read Welcome to the Era of Physical Evals →