multimodal_gen runtime.
This section groups dedicated action-policy recipes separately from media recipes so robot deployments can document their own input schemas, policy endpoints, cache behavior, and control-loop performance targets. See the action API for shared endpoints and Cosmos3 for world-model action generation.
VLA (Vision-Language-Action) Models
Overview
Robot action generation with SGLang Diffusion and Vision-Language-Action policies.
SGLang Diffusion supports action generation alongside image, video, and audio generation. Vision-Language-Action policies map camera observations, language instructions, and robot state into continuous action chunks using the same
