| --- |
| language: en |
| pipeline_tag: robotics |
| tags: |
| - openwam |
| - domino |
| - aloha-agilex |
| - video-action-model |
| - safetensors |
| --- |
| # OpenWAM DOMINO Foundation — 10,000 steps |
|
|
| Foundation fine-tuning of **OpenWAM-Alpha-Pretrain-Foundation-Model** on all **35 DOMINO tasks**, using **Aloha-AgileX / clean / level1** demonstrations. This is a joint video/action model, not the Piper Stage2 residual model. |
|
|
| ## Training |
| - Official OpenWAM source commit: `3dfc9a7d9230f0e1382e6b22639278f1a86e5eb5` (with DOMINO data preparation and cached T5 adaptations). |
| - Initialization: `OpenWAM-Alpha-Pretrain-Foundation-Model`, step 154000. |
| - Data: 1,750 episodes, 318,976 frames and 317,226 training windows; 50 episodes per task. |
| - Training: 10,000 optimizer steps; 6 GPUs, batch size 32 per GPU (global batch 192). |
| - Frozen T5 and VAE; joint video/action training. |
| - Input: three RGB cameras, L-shaped canvas 384 x 320; 33 source frames with stride 4, producing 9 video frames and 32 action targets. |
| - Representation: end-effector position + rotation-6D + gripper, unified state/action width 80; raw 20D dimensions map to `0-9,34-43`. |
| - Min-max normalization; exact architecture and representation are saved in `config.yaml`. |
|
|
| ## Files |
| - `checkpoint_step_10000.safetensors`: final checkpoint with model, VAE and text encoder weights. |
| - `config.yaml`, `normalization_stats.npy`, `tokenizer/`: required self-contained deployment assets. |
| - `SHA256SUMS`: integrity manifest for the checkpoint and deployment assets. |
| - `release_metadata.json`: training provenance and checkpoint verification. |
|
|
| ## Evaluation |
| DOMINO simulation evaluation is being configured separately for Aloha-AgileX / clean / level1 with unseen instructions. No simulation success-rate claim is made in this release. Training losses are not benchmark scores. |
|
|
| The saved training paths are provenance only; OpenWAM deployment reconstructs model components from the embedded specifications and restores all weights from this checkpoint. On 24 GB GPUs, keep the frozen T5 text encoder on CPU or use precomputed prompt embeddings to reserve GPU memory for the video/action network. |
|
|
| Respect the licenses and usage terms of the original OpenWAM models, DOMINO data, simulator and assets. |
|
|
