World Action Models
World action models are robot policies built on video generation: a video diffusion transformer predicts the next frames of the robot’s cameras from the task description and the current observation, and an action expert turns those predictions into the next chunk of actions. Fine-tuning teaches both halves at once: the trunk learns the new scene and controller, the action expert learns the new action space.
Run World Action Models with NeMo AutoModel
Use a source checkout of NeMo AutoModel so that the package and its example recipes stay on one revision:
For other installation options, see our NeMo AutoModel Installation Guide.
The recipes load the frozen encoders of each model (for Praxis-1, the Wan2.2 VAE and the Qwen3.5 text encoder) through the model builder’s reference package, which must be on the Python path. The model cards below name it.
Supported Models
This table combines recipe-backed checkpoints with documented model families. Dates show when the current checkpoint first appeared in a recipe, or when a documentation-only family page was added. See the combined model support log for recipe-backed checkpoints of every model type.
Fine-Tune World Action Models
The checked-in recipes fine-tune all weights with FSDP2 and expert parallelism on one 8-GPU node, or train LoRA adapters on the trunk and the action expert. The Praxis-1 LIBERO tutorial walks through converting a benchmark’s demonstrations, training, and evaluating before and after.