The Problem

State-of-the-art vision-language models can describe what they see, but they cannot reliably act in the world. The bottleneck is a structural gap in training data across space, physics, and embodied action. PRISM closes that gap.

Key Findings

• • Domain-specific SFT beats general pretraining, reducing average error by 66.6%.
• • Multi-view training is mutually reinforcing; exocentric supervision improves ego performance.
• • 60% of the data captures 95% of the total performance ceiling.

Impact Metrics