Robotics is unsettled: world models are taking ground from VLA — resolves 2027-07-14

The debate: there are two routes to a robot's brain, VLA, which puts seeing, reading instructions and acting into a single model, and world models, which first learn to predict how the environment changes and then use that to control the robot.

Why it matters: VLA leads in product deployment today, while world models can pretrain on internet-scale video, orders of magnitude more data than robot-specific action data, so the winner decides which foundation robot companies bet on.

Who is on each side: Gemini Robotics, Figure and the pi series take the VLA route and lead in deployment, while Cosmos Policy post-trained a video generation model directly into a control policy and beat the best fine-tuned VLA on the same data.

SecondSource’s reading: whether robots should run on VLA, a model that unifies vision, language and action, or on a world model, one that first learns to predict how the environment changes, is not settled.

VLA leads in product deployment. But Cosmos Policy post-trained a video generation model directly into a robot control policy and beat the best fine-tuned VLA on the same data, a sign that world-model foundations are starting to take ground from VLA.

For now this is a single case, and what matters is whether other independent teams can get the same result.

Bruce et al., Google DeepMind, arXiv
“We introduce Genie, the first generative…”

Over 12 months: two or more independent teams reproducing a video or world model post-trained directly into a policy that beats a VLA on the same data confirms the world-model side. No reproduction leaves VLA's default position intact.

What would overturn it

Two or more independent reproductions void VLA-specific architectures as the default. No one reproducing world-model policies downgrades the claim that the boundary is blurring.