Axis Weekly We've been experimenting with how to digest data better. After almost a year of building the data flywheel, the next step is a model-guided data engine: use foundation-model performance as feedback to decide what to keep, what to drop, and where the next batch should be produced. We already made large-scale data generation and collection possible. The remaining bottleneck is knowing what the current model still lacks, so we can filter existing data and aim the next collection round instead of sampling from a fixed distribution. Thesis article on this soon. Key updates: • New Dataset & Real2Sim2Real: Pipeline for bringing real objects and scenes into simulation, with hardware validation. Focus on cross-object and cross-scene skills on concrete families first. • Active curation: Moving from human/VLM/script checks to model-performance-driven selection, so training budget goes to trajectories that still teach something. • Verification & new tasks: Hardened anti-cheating checks, dual-arm support with ARX5, and new put-in/take-out basket tasks. Details below 👇