Evaluate. Collect. Train.
Foundation models for the physical world start with knowing what works. B2A is building in that order.
Evaluate — live. Independent, source-checked evaluation of physical AI models. AI Model Performance tracks 15 models on six criteria: success, robustness, instruction, recovery, speed and interventions. 38 of 90 values are published; 52 are still owed. AI assistants can query it directly through its MCP server.
Collect — coming. Data from robots doing production work: cameras, joint states and outcomes, recorded in the cell rather than staged in a lab. We’re lining up the first data partners.
Train — later. Models for the physical world, trained on that data and measured the same way as everything on the scorecard.
The order matters. A model you can’t measure can’t be trusted on a line, and data that isn’t tied to outcomes can’t train one that can.