ChrisF(✱,✱)(@chris_anm01):What V1 proved: An end-to-end sim-to-real pipeline works. Engineered diversity beats raw volume (79.4 on LIBERO-Plus), and transfers directly to real physical robots. The bottleneck: V1 is unidirectional. Once deployed, the model can't recover from its own failures — because the data flow stops after training. The V2 shift: Close the loop. Human-Gated DAgger turns model errors into targeted data collection. Axis is no longer a pipeline — it's a compounding, human-in-the-loop data engine.

2026.07.09 15:03

What V1 proved: An end-to-end sim-to-real pipeline works. Engineered diversity beats raw volume (79.4 on LIBERO-Plus), and transfers directly to real physical robots. The bottleneck: V1 is unidirectional. Once deployed, the model can't recover from its own failures — because the data flow stops after training. The V2 shift: Close the loop. Human-Gated DAgger turns model errors into targeted data collection. Axis is no longer a pipeline — it's a compounding, human-in-the-loop data engine.