We wrote a blog about our key finding from Axis Sim Dataset V1: given sufficiently large and diverse, in-the-wild data, behavioral cloning converges toward a robust policy — even when individual demonstrations are noisy. Think of it like many people pulling a heavy load with ropes. Each person pulls at a slightly different angle, but everyone pulls forward. When there are enough people and the goal is shared, the sideways forces weaken each other while the forward forces add up, and the load moves steadily forward. Every valid trajectory works the same way. It mixes the actions that complete the task with the operator's own deviations — jitter, hesitation, suboptimal paths. These deviations aren't systematic; they point in different directions and weaken each other at scale. The task-completing actions, all pointing the same way, get reinforced. We validated this in our paper. 50K+ community-driven trajectories from Axis Hub lifted π0.5 on LIBERO-Plus from 83.9% to 88.8%. The point isn't how clean each trajectory is — it's how broad the distribution is. Base capability comes from scale; the last mile comes from DAgger — the community continuously correcting the gaps the model reveals. And that's what Axis V2 is about. Read the full blog: