| --- |
| license: other |
| library_name: pytorch |
| pipeline_tag: robotics |
| datasets: [XDOF/ABC-130k] |
| tags: [long-wam, yam, abc-130k, robotics] |
| --- |
| |
| # Long-WAM YAM |
|
|
| A YAM bimanual robot pretraining model trained on **ABC 130K**, with a |
| LongLive 2.0 Robot-S video backbone. It predicts actions after imagining future |
| video latents. |
|
|
| ## Download |
|
|
| ```bash |
| hf download Efficient-Large-Model/Long-WAM-YAM --local-dir ./weights/Long-WAM-YAM |
| ``` |
|
|
| ## Inference |
|
|
| Load `model.pt` with the matching Long-WAM runtime, `config.yaml` and |
| `dataset_stats.json`. Use 4 video denoising steps to sigma 0.9, followed by |
| 10 action denoising steps. |
|
|
| The three RGB views use a fixed **384 × 320** layout: |
|
|
| ```text |
| +---------------------------------+ |
| | Top view: 256 × 320 | |
| +----------------+----------------+ |
| | Left wrist | Right wrist | |
| | 128 × 160 | 128 × 160 | |
| +----------------+----------------+ |
| ``` |
|
|
| P48 covers **1.6 seconds at 30 Hz**. Sample history every four control steps: |
| 13 images become four clean latent frames, followed by two imagined latents. |
| Predict a 32-step action chunk, approximately 1.07 seconds. |
|
|
| State and action are 14-D: left arm's six joints and gripper, then right arm's |
| six joints and gripper. Use the supplied z-score statistics to normalize state |
| and denormalize predicted actions; RGB inputs use `[-1, 1]`. |
|
|
| The matching runtime and base-model VAE/text assets are required separately. |
| This is a pretraining checkpoint: task-specific adaptation and controlled |
| closed-loop validation are required before real-robot use. |
|
|
