WE-Dream Pretraining Models
Collection
Action-free, hand-pose, and latent-action pretrained models for robot post-training. β’ 5 items β’ Updated
This action-free WE-Dream checkpoint was pretrained for 50,000 steps on approximately 2,000 hours of EgoDex and EgoVerse egocentric video, using video prediction without action supervision. It initializes robot post-training and provides a comparison with hand-pose and latent-action pretraining.
| Property | Value |
|---|---|
| Prediction horizon | 24 steps |
| Routed-attention probability | 0.5 |
| Weights | Safetensors, 19 shards, approximately 91.69 GB |
| Saved precision | Primarily FP32 |
Follow the installation instructions to set up the training environment and download the base model files. Accept the access terms on this page, then download:
hf download we-dream/pretrain-action-free-2kh \
--local-dir checkpoints/we-dream-action-free-2kh
Set PRETRAINED_MODEL_PATH=checkpoints/we-dream-action-free-2kh when following the post-training guide.