What's this about?
The day's top paper at 92 upvotes. Vision-language-action models are now the dominant approach for general-purpose robots and do well on long tasks in structured settings — but whether they gain genuinely new capabilities with scale is open. GigaBrain-0.7 attacks that with a three-system architecture.
What exactly is a "vision-language-action" model?
One model that takes in what the robot sees, plus an instruction in ordinary language, and outputs motor commands directly. Before this approach you built separate perception, planning and control systems and glued them together. Doing it end-to-end means the model can generalise to instructions and objects it was never explicitly programmed for — which is also why measuring what it actually learned is difficult.
What's the community saying?
The word people focus on is *emergent* — capabilities that appear at scale rather than being trained for. It is a contested claim in language models and even harder to establish in robotics, where evaluation is expensive and physical. The three-system split is the paper's structural bet on getting there.