Evaluating Gemini Robotics policies in the Veo world simulator

The Gemini Robotics team proposes a generative evaluation system based on the Veo video foundation model, for evaluating the nominal performance, out-of-distribution (OOD) generalization, and physical and semantic safety of robot visuomotor policies. The system is optimized for robot action-conditioned control and multi-view consistency, and combines generative image editing with multi-view completion to synthesize realistic scene variations containing new interacting objects, new visual backgrounds, and new distractor objects. The paper states that the system preserves the video model's foundational capabilities, can accurately predict the relative performance of different policies under nominal and out-of-distribution conditions, and exposes policy behaviors that violate physical or semantic safety constraints through red-team testing. The team validated these capabilities through more than 1600 real-world evaluations, involving eight Gemini Robotics policy checkpoints and five tasks on a dual-arm robotic arm.