Building Worlds That Train Robots

July 28, 2026

World Labs shares early results from its Real-to-Sim-to-Real (R2S2R) engine acquired via SceniX (joined July 21, 2026).

One physical task generates many controllable simulation variants. Policies trained entirely in simulation transferred to diverse real robots including ALOHA, preserving model ranking on cube-handoff benchmarks.

Key claims: zero real-world training data for complex manipulation; simulated evaluation predicts hardware policy performance; policies ran for hours on physical robots.

No public SDK or open-source release announced as of publication date.