I counted 47 new robotics foundation models announced in the last 12 months.Papers. Demos. Benchmark results. Twitter threads with cherry-picked clips.How many are running live in a warehouse right now, picking real customer orders?You can count them on one hand.The gap between "works in the lab" and "works at 2am on a Friday in a fulfillment center outside Rotterdam" is not a scaling problem. It's a different problem entirely.Lab conditions forgive you. Fixed lighting. Clean bins. Objects place ...

Sereact

Nobody's automation pilot fails. The pilot always works.It's month 4 at higher volume when the numbers stop matching the deck. A new SKU shape enters the returns stream. Intervention rates quietly double. The gap between "it can pick" and "it picks reliably at production scale" becomes the entire business case.Reactive policies hit this wall because they have no model of what happens next. They optimize for the immediate action given the current observation. When a grasp fails, they retry with t ...

Sereact

A VLA policy picks the best next action given what it sees right now. Usually that works. But put it in a cluttered returns bin with unknown object poses and a task that takes 15 steps, and a perfectly reasonable early grasp can lead to a jam or collision on step 9 that the policy never modeled.Cortex 2.0 adds a planning loop without replacing the execution stack.The world model encodes the current observation into a latent state, then rolls out multiple candidate trajectories in visual latent s ...

Sereact