minimal card (attribution+license only)

X-Actor-GR00T/README.md

@@ -8,32 +8,10 @@ tags:
   - robotics
   - manipulation
   - vision-language-action
-  - multi-view
-  - viewpoint-robustness
-  - libero
 ---
 
 # X-Actor-GR00T
 
-**X-Actor** method applied to the **GR00T-N1.5-3B** vision-language-action backbone, fine-tuned on LIBERO for **camera-viewpoint–robust** robotic manipulation.
+X-Actor checkpoint built on the **GR00T-N1.5-3B** backbone, fine-tuned on LIBERO.
 
-X-Actor adds an object-aware multi-view foresight representation to the policy and injects it into the action expert through a zero-initialized adapter. At init the model is equivalent to the bare backbone; training then teaches it to stay accurate when the agent-view camera is rotated/shifted — the regime where standard VLAs collapse.
-
-## Base model & license (please read)
-This checkpoint is a **fine-tune of NVIDIA GR00T-N1.5-3B and embeds its weights**. Your use of these weights is governed by **NVIDIA's GR00T-N1.5 model license** (link above). A frozen **DINOv2-with-registers** (Meta) encoder is used as an auxiliary visual feature extractor. X-Actor is the fine-tuning/representation method; the foundation model is NVIDIA's.
-
-## What it shows (LIBERO-Spatial, agent-view camera perturbation)
-Like-for-like, same GR00T-N1.5 backbone, **X-Actor vs. the bare backbone**:
-
-| perturbation | X-Actor + GR00T | bare GR00T-N1.5 |
-|---|---|---|
-| clean | **96.0** | 91.5 |
-| yaw 15° | **56.7** | 13.3 |
-| yaw 30° | **45.5** | 0.0 |
-| yaw 45° | **33.0** | 0.0 |
-| pitch 30° | **33.0** | 3.3 |
-
-The bare backbone drops to ~0% under ≥30° viewpoint rotation; X-Actor recovers a large fraction of the success rate while keeping clean performance.
-
-## Intended use
-Research checkpoint for studying viewpoint-robust manipulation policies on LIBERO. Not a production controller.
+**Base model & license:** this checkpoint embeds **NVIDIA GR00T-N1.5-3B** weights; use is governed by NVIDIA's GR00T-N1.5 model license (link above). DINOv2-with-registers (Meta) is used as a frozen feature extractor.