vllm_omni.diffusion.models.dreamzero.transform.droid ¶
DROID dataset transform.
DROID uses 1-indexed exterior cameras, 3 views total (OXE_DROID embodiment). Stitching layout (same as RoboArena — both are OXE_DROID): ┌─────────────────────────┐ │ wrist (2x width) │ ← pixel-repeat along width ├────────────┬────────────┤ │ left ext │ right ext │ └────────────┴────────────┘
DroidTransform ¶
Bases: RobotPolicyTransform
Transform for DROID dataset (OXE_DROID embodiment).
DROID observation keys (1-indexed exterior cameras): observation/exterior_image_1_left → left exterior observation/exterior_image_2_left → right exterior observation/wrist_image_left → wrist
IMAGE_KEY_MAP class-attribute instance-attribute ¶
IMAGE_KEY_MAP = {
"observation/exterior_image_1_left": "images/exterior_0",
"observation/exterior_image_2_left": "images/exterior_1",
"observation/wrist_image_left": "images/wrist",
}