| # InternW0 RoboTwin-clean checkpoint provenance |
|
|
| ## Selected artifact |
|
|
| `InternW0-RoboTwin-clean.pt` is the model-only checkpoint selected for the |
| RoboTwin 2.0 Clean2Random release. It is derived from the historical run |
| `w0-cln-cj02-s2-64g-dh-0919r5` at step 10,545. |
|
|
| The historical source implementation was `ahawam`; the public runtime is `w0`. |
| The resolved numerical training contract is preserved in |
| `source_training_config.yaml` with private filesystem paths replaced by portable |
| environment variables. |
|
|
| ## Lineage |
|
|
| 1. Heterogeneous InternW0 pretraining at global step 30,000. |
| 2. RoboTwin target-domain initialization with `new_domain` action/proprio |
| interfaces, seed 42. |
| 3. Clean-demonstration posttraining run |
| `w0-cln-p30k-cj02-s2-64g-0918r3`, resumed at step 5,000. |
| 4. Continued best Clean2Random recipe |
| `w0-cln-cj02-s2-64g-dh-0919r5`, selected at step 10,545. |
|
|
| The checkpoint embeds `behavior_config`, `training_recipe`, |
| `posttrain_statistics`, and `posttrain_provenance`. These fields are also |
| exported verbatim to `checkpoint_metadata.json`. |
|
|
| ## Training contract |
|
|
| - Clean demonstrations only. |
| - RoboTwin native joint-space representation. |
| - 25 Hz sampling (`sampling_stride=2` from 50 Hz). |
| - Every valid native window. |
| - ColorJitter: brightness 0.2, contrast 0.2, saturation 0.2, hue 0. |
| - Offset 0. |
| - Batch size 12 per process. |
| - 15 epochs. |
| - Learning rate 3e-5, cosine schedule. |
| - Seed 42. |
| - 37-dimensional canonical action/proprio interface. |
| - 14 valid ALOHA joint/gripper slots: |
| `[0, 1, 2, 3, 4, 5, 16, 17, 18, 19, 20, 21, 22, 33]`. |
| - Action horizon 64, chunk size 16, video plan horizon 96, video/action stride 8. |
| - K=8 target soft prompts. |
| - Native absolute joint commands at the interval end. |
|
|
| The matching normalization file is `normalization.json`. |
|
|
| ## Evaluation reference |
|
|
| The source resolved evaluation reference is preserved in |
| `source_evaluation_config.yaml`. It uses the fixed-period RoboTwin adapter at |
| 25 Hz, 10 inference steps, a 64-action horizon, four action chunks per video |
| prefill, and 15 episodes per task. |
|
|
| The paper-reported RoboTwin 2.0 Clean2Random results are: |
|
|
| | Split | Success rate | |
| |---|---:| |
| | Clean | 83.20% | |
| | Randomized (unseen during training) | 68.00% | |
| | Average | 75.60% | |
|
|
| ## Excluded from this repository |
|
|
| The checkpoint package does not include optimizer state, distributed training |
| state, logs, RoboTwin data, Wan2.2 files, DINOv3 weights, text caches, or |
| simulation checkouts. |
|
|
