ViGAR weights

Selected RoboTwin inference weights for ViGAR: an action policy and an image subgoal planner.

Component	Folder	Selection
Action policy	policy/robotwin_c2r	three views, native 49-D, 50k EMA
Subgoal planner	subgoal_planner/robotwin_subgoal_planner	regular weights

The weights use PyTorch Distributed Checkpoint (DCP). Download all .distcp shards and .metadata. See the installation and evaluation guide for the runtime setup.

Download

hf download DAGroup-PKU/ViGAR --local-dir /path/to/vigar-weights

Then set in the code repo's .env:

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

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 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.

Policy training data

The segmented dataset 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:

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.

License

Cosmos-derived model materials are under NVIDIA OpenMDW-1.1 (see LICENSE and NOTICE). The Wan2.2 VAE and other auxiliary components keep their upstream terms under licenses/.