| --- |
| license: apache-2.0 |
| library_name: pytorch |
| tags: |
| - robotics |
| - robot-manipulation |
| - world-model |
| - imitation-learning |
| - video-policy |
| - optical-flow |
| - domino |
| - dynamicwam |
| --- |
| |
| # DynamicWAM |
|
|
| Official DynamicWAM checkpoints for DOMINO manipulation. This repository is |
| organized by paper-facing model variant; each checkpoint is paired with its |
| action-normalization statistics and immutable configuration snapshots. |
|
|
| ## Released variants |
|
|
| | Paper name | Checkpoint | Motion conditioning | DOMINO Level 1 | |
| |---|---|---|---:| |
| | **DynamicWAM (w/o motion)** | [`checkpoints/dynamicwam_wo_motion/model.pt`](checkpoints/dynamicwam_wo_motion/model.pt) | Four history Flow-RGB intervals; **no explicit absolute-motion descriptor tokens** | 953/3,500, SR 27.23%, MS 41.6244 | |
| | **DynamicWAM** | [`checkpoints/dynamicwam/model.pt`](checkpoints/dynamicwam/model.pt) | History Flow RGB + exact-time absolute-motion descriptors | 1,337/3,500, SR 38.20%, MS 53.16 | |
|
|
| “w/o motion” refers specifically to removing the explicit numeric |
| absolute-motion/kinematic descriptor branch. It does **not** mean that optical |
| flow is removed: the variant still consumes history Flow RGB. |
|
|
| The reported numbers are the archived 35-task × 100-episode DOMINO Level 1 |
| system results. They are not a newly rerun benchmark and should not be |
| interpreted as a matched one-variable causal estimate beyond the named model |
| variants. |
|
|
| ## Repository layout |
|
|
| ```text |
| checkpoints/ |
| ├── dynamicwam_wo_motion/ |
| │ ├── model.pt |
| │ └── action_stats.json |
| └── dynamicwam/ |
| ├── model.pt |
| └── action_stats.json |
| configs/ |
| ├── dynamicwam_wo_motion/ |
| │ ├── stage1_video.yaml |
| │ ├── stage2_action.yaml |
| │ ├── stage3_joint.yaml |
| │ └── deploy.yaml |
| └── dynamicwam/ |
| └── absolute_motion_v2.yaml |
| MODEL_MANIFEST.json |
| SHA256SUMS |
| ``` |
|
|
| The `dynamicwam_wo_motion` YAML files are exact historical configuration |
| snapshots and therefore retain their original cluster paths. Remap those paths |
| to your local checkout and downloaded assets before training or evaluation. |
|
|
| ## Download |
|
|
| ```bash |
| hf download KhalilGao/DynamicWAM \ |
| --include "checkpoints/dynamicwam_wo_motion/*" \ |
| --include "configs/dynamicwam_wo_motion/*" \ |
| --local-dir external/DynamicWAM |
| ``` |
|
|
| The matching packed training corpus is published at |
| [`KhalilGao/DynamicWAM-data`](https://huggingface.co/datasets/KhalilGao/DynamicWAM-data) |
| under `dynamicwam_wo_motion/`. |
|
|
| ## Checkpoint identity |
|
|
| | Variant | Global step | Size (bytes) | SHA-256 | |
| |---|---:|---:|---| |
| | DynamicWAM (w/o motion) | 40,000 | 1,976,638,827 | `fa88a1f33f2205db5a4534e09bb02d1bee9dea7b2049299ac68932a34de82f03` | |
| | DynamicWAM | 40,000 | 1,977,857,939 | `7c0dfc44a785ea1f6bd1f833f09dcadc2e470dadb1ba5508fa98918e147671d7` | |
|
|
| Both are native PyTorch checkpoint dictionaries. As with any pickle-backed |
| PyTorch artifact, load them only from a trusted revision and verify the hashes |
| in `SHA256SUMS` first. |
|
|
| ## External assets |
|
|
| The checkpoints do not bundle Wan2.2-TI2V-5B, its tokenizer/VAE, the DOMINO |
| runtime, CuRobo, or simulator assets. See the |
| [`DynamicWAM` source repository](https://github.com/Autumn1337/DynamicWAM) |
| for environment and evaluation setup. |
|
|
| ## License |
|
|
| Project-owned release material is provided under Apache-2.0. Upstream assets |
| retain their own licenses; this repository does not relicense or redistribute |
| the external runtime dependencies listed above. |
|
|
