Move model weights under weights/ and document the Cosmos3-Nano base checkpoint

policy/ -> weights/policy/, subgoal_planner/ -> weights/subgoal_planner/. README, NOTICE and release_config.json use the new paths.

NOTICE

@@ -2,13 +2,16 @@ ViGAR model materials
 
 This repository contains the ViGAR action-policy and subgoal-planner DCP
 checkpoints, along with the policy bundle's VAE, tokenizer and configuration
-assets. ViGAR provides fine-tuned RoboTwin checkpoints and portable inference
-bundle metadata.
+assets, and a DCP conversion of the Cosmos3-Nano base checkpoint. ViGAR provides
+fine-tuned RoboTwin checkpoints and portable inference bundle metadata.
 
 Cosmos3-Nano
 Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
 The policy and subgoal-planner checkpoints derive from Cosmos3-Nano model
-materials. These materials are distributed under OpenMDW-1.1; see LICENSE.
+materials. weights/Cosmos3-Nano-dcp is Cosmos3-Nano (revision
+fea6e03ac3d7884b4105ed8ee79fc480fca70965) converted to PyTorch Distributed
+Checkpoint format without other changes. These materials are distributed under
+OpenMDW-1.1; see LICENSE.
 Upstream: https://huggingface.co/nvidia/Cosmos3-Nano
 
 Wan2.2 VAE

README.md

@@ -12,19 +12,26 @@ tags:
 
 # ViGAR weights
 
-Selected RoboTwin inference weights for [ViGAR](https://github.com/DAGroup-PKU/ViGAR):
-an action policy and an image subgoal planner.
+RoboTwin weights and policy training data for [ViGAR](https://github.com/DAGroup-PKU/ViGAR):
+an action policy, an image subgoal planner, the Cosmos3-Nano base checkpoint used
+to train the policy, and the segmented training dataset.
 
 | Component | Folder | Selection |
 |---|---|---|
-| Action policy | `policy/robotwin_c2r` | three views, native 49-D, 50k EMA |
-| Subgoal planner | `subgoal_planner/robotwin_subgoal_planner` | regular weights |
+| Action policy | `weights/policy/robotwin_c2r` | three views, native 49-D, 50k EMA |
+| Subgoal planner | `weights/subgoal_planner/robotwin_subgoal_planner` | regular weights |
+| Base checkpoint | `weights/Cosmos3-Nano-dcp` | Cosmos3-Nano, policy training start point |
+| Policy training data | `data/robotwin_segmented_2500` | 2,500 episodes, 3,691 goals |
 
-The weights use PyTorch Distributed Checkpoint (DCP). Download all `.distcp`
-shards and `.metadata`. See the [installation and evaluation guide](https://github.com/DAGroup-PKU/ViGAR#installation) for the runtime setup.
+The weights use PyTorch Distributed Checkpoint (DCP). Each checkpoint `model/`
+folder holds its `.distcp` shards and `.metadata`. See the
+[installation and evaluation guide](https://github.com/DAGroup-PKU/ViGAR#installation)
+for the runtime setup.
 
 ## Download
 
+Download the whole repository (about 96 GB) once:
+
 ```bash
 hf download DAGroup-PKU/ViGAR --local-dir /path/to/vigar-weights
 ```
@@ -32,32 +39,40 @@ hf download DAGroup-PKU/ViGAR --local-dir /path/to/vigar-weights
 Then set in the code repo's `.env`:
 
 ```bash
-VIGAR_POLICY_CHECKPOINT=/path/to/vigar-weights/policy/robotwin_c2r
-SUBGOAL_PLANNER_CHECKPOINT=/path/to/vigar-weights/subgoal_planner/robotwin_subgoal_planner
-WAN_VAE_PATH=/path/to/vigar-weights/policy/robotwin_c2r/vae/Wan2.2_VAE.pth
-QWEN_TOKENIZER_PATH=/path/to/vigar-weights/policy/robotwin_c2r/text_tokenizer
+VIGAR_POLICY_CHECKPOINT=/path/to/vigar-weights/weights/policy/robotwin_c2r
+SUBGOAL_PLANNER_CHECKPOINT=/path/to/vigar-weights/weights/subgoal_planner/robotwin_subgoal_planner
+WAN_VAE_PATH=/path/to/vigar-weights/weights/policy/robotwin_c2r/vae/Wan2.2_VAE.pth
+QWEN_TOKENIZER_PATH=/path/to/vigar-weights/weights/policy/robotwin_c2r/text_tokenizer
+BASE_CHECKPOINT_PATH=/path/to/vigar-weights/weights/Cosmos3-Nano-dcp
+VIGAR_DATASET_MANIFEST=/path/to/vigar-weights/data/robotwin_segmented_2500/dataset.json
+VIGAR_GENERATED_GOAL_CACHE=/path/to/vigar-weights/data/robotwin_segmented_2500/goal_cache
 ```
 
+Evaluation uses the policy and planner. Policy training uses the base checkpoint,
+the dataset and the policy bundle's tokenizer and Wan VAE.
+
+## Policy and planner
+
 The policy bundle carries its architecture config, tokenizer, Wan VAE, recipe and
-normalizer. The planner can reuse that bundle's tokenizer/VAE as described in its
+normalizer. The planner reuses that bundle's tokenizer/VAE as described in its
 `inference_config.json`. Inference uses current-observation Strict Sync — policy 10
 steps / guidance 1 / shift 2 (predict 48, execute 32); planner 35 steps / guidance
 2.5 / shift 5 at RGB 384×320.
 
+## Base checkpoint
+
+`weights/Cosmos3-Nano-dcp` is [nvidia/Cosmos3-Nano](https://huggingface.co/nvidia/Cosmos3-Nano)
+(revision `fea6e03ac3d7884b4105ed8ee79fc480fca70965`) converted to DCP with
+`cosmos_framework.scripts.convert_model_to_dcp`. Its `model/` folder has six shards
+of about 5 GB each. Policy training loads it through `BASE_CHECKPOINT_PATH`; it is
+not needed for evaluation.
+
 ## Policy training data
 
 The [segmented dataset](data/robotwin_segmented_2500) contains 2,500 RoboTwin
 training episodes: 19 multi-stage tasks and 31 final-goal tasks, with 3,691 selected
-generated goals. Download it from this repository:
-
-```bash
-hf download DAGroup-PKU/ViGAR --include "data/robotwin_segmented_2500/**" --local-dir /path/to/vigar-weights
-```
-
-Use `data/robotwin_segmented_2500/dataset.json` as `VIGAR_DATASET_MANIFEST` and
-`data/robotwin_segmented_2500/goal_cache` as `VIGAR_GENERATED_GOAL_CACHE`, with the
-absolute paths of your download. Dataset licensing and attribution are in its
-`LICENSE` and `NOTICE` files.
+generated goals. Dataset licensing and attribution are in its `LICENSE` and
+`NOTICE` files.
 
 ## License
 

release_config.json

@@ -1,9 +1,10 @@
 {
   "repository": "DAGroup-PKU/ViGAR",
-  "policy": "policy/robotwin_c2r",
+  "policy": "weights/policy/robotwin_c2r",
   "policy_weights": "ema",
   "policy_name": "robotwin_c2r",
-  "subgoal_planner": "subgoal_planner/robotwin_subgoal_planner",
+  "subgoal_planner": "weights/subgoal_planner/robotwin_subgoal_planner",
   "subgoal_planner_weights": "regular",
+  "base_checkpoint": "weights/Cosmos3-Nano-dcp",
   "policy_dataset": "data/robotwin_segmented_2500"
 }