Upload README.md with huggingface_hub

README.md

@@ -1,18 +1,18 @@
 ---
 license: other
-library_name: pytorch
+library_name: cosmos-policy
 pipeline_tag: robotics
 ---
 
 # Cosmos Policy MetaWorld Predict2 2B
 
-Artifacts for the FlowDAgger OSS Cosmos/MetaWorld assembly example.
+FlowDAgger OSS release artifact for the Cosmos 2B MetaWorld backend.
 
-This bundle contains:
-- `consolidated.pt`: Cosmos Policy Predict2 2B checkpoint finetuned on MetaWorld.
-- `dataset_statistics.json`: normalization statistics used by the policy wrapper.
-- `t5_embeddings.pkl`: cached text embeddings for the MetaWorld prompts.
+Files:
+- `consolidated.pt`: Cosmos Policy Predict2 2B checkpoint finetuned on MetaWorld data for the release recipe.
+- `dataset_statistics.json`: action/proprio normalization statistics.
+- `t5_embeddings.pkl`: cached text embeddings for the MetaWorld prompts used by the example.
 
-The released FlowDAgger recipe targets MetaWorld `assembly-v3`. In the July 2026 reproduction, the frozen base policy reached about 52% success over 25 evaluation episodes and FlowDAgger full-mode steering reached about 84-88%.
+The release recipe targets MetaWorld `assembly-v3` with prompt `Pick up a nut and place it onto a peg`. This checkpoint intentionally leaves headroom: the frozen base policy is about 52% SR over 25 eval episodes in the reference run, and FlowDAgger full-mode steering reaches about 84-88%.
 
-The checkpoint expects the MetaWorld gripper proprio convention `[g, -g]`; use `--gripper_qpos_mode opposed` in the OSS launcher.
+The checkpoint expects MetaWorld gripper proprio encoded as `[g, g]`; use `--gripper_qpos_mode duplicated` with the FlowDAgger Cosmos runner.