The dataset viewer is not available because its heuristics could not detect any supported data files. You can try uploading some data files, or configuring the data files location manually.
RoboRender anchor pool (2048)
Teacher-rendered RGB anchors for FT-FIXED conditioning of the Image3F 3-view image model.
Each anchor_NNNNN.npz holds two frames per view from a 50-step teacher video-model clip:
| key | shape | note |
|---|---|---|
gen__head_color |
(2, 240, 416, 3) uint8 | frames [0, 40] |
gen__hand_left |
(2, 240, 416, 3) uint8 | |
gen__hand_right |
(2, 240, 416, 3) uint8 | |
frames |
(2,) int64 | [0, 40] |
prompt |
str | scene description used to render it |
src |
str | source depth chunk |
Index 0 is the episode-start triplet, which production uses (videogen.anchor_index: 0)
and which matches the adapter's fixed-lag-256 training contract. Index 40 is a
content-rich alternative: frame 0 leaves the wrist references nearly empty at episode start,
and a later frame measurably improves colour fidelity in exactly those views. Slightly
off-contract but stable.
Prompts come from a combinatorial generator (~29M combinations over walls, floors, tables, objects, containers and lighting), so the anchors are the domain randomisation for the generated-observation path. The trainer reads the prompt from the anchor file.
Resolution is 240x416 per view — the renderer's native size, not the policy's 108x192. These are conditioning inputs to Image3F, not observations.
Rendered with teacher LoRA ...-droidnew_continue_lora/step-5294.safetensors
(md5 20e940246337b3a73156d647cef1a82d).
- Downloads last month
- 31