Skip to content

vllm_omni.diffusion.models.dreamzero.transform.roboarena

RoboArena dataset transform.

RoboArena uses 0-indexed exterior cameras, 3 views total (OXE_DROID embodiment). Same stitching layout as DROID — both map to OXE_DROID in DreamZero.

RoboArenaTransform

Bases: DroidTransform

Transform for RoboArena dataset.

Same embodiment as DROID (OXE_DROID), same stitching and template. Only difference: 0-indexed exterior camera keys.

RoboArena observation keys (0-indexed): observation/exterior_image_0_left → left exterior observation/exterior_image_1_left → right exterior observation/wrist_image_left → wrist

IMAGE_KEY_MAP class-attribute instance-attribute 

IMAGE_KEY_MAP = {
    "observation/exterior_image_0_left": "images/exterior_0",
    "observation/exterior_image_1_left": "images/exterior_1",
    "observation/wrist_image_left": "images/wrist",
}