praxis-1-droid | NVIDIA NeMo AutoModel

Praxis-1 is Runway’s 27B world-action model: a sparse mixture-of-experts video diffusion transformer (about 3B active parameters) that predicts the next frames of the robot’s cameras, and an action expert that reads those predictions and outputs the next chunk of end-effector actions. NeMo AutoModel has a checked-in recipe that fine-tunes all of its weights on one node, and a tutorial that takes the single-arm checkpoint from 0 to 90 % success on a LIBERO task suite.

Fine-Tune Praxis-1

Follow the installation instructions from a source checkout. The recipe also needs the reference praxis package of the model builder on the Python path: it provides the Wan2.2 VAE and the Qwen3.5 text encoder that the policy is trained with, and the closed-loop simulation client the tutorial evaluates with.

Choose a Workflow

Goal	Start Here
Fine-tune all weights on one 8-GPU node	Use the recipe configuration.
Convert LIBERO demonstrations, train, and evaluate before / after	Follow the LIBERO tutorial.
Train adapters instead of all weights	Use the LoRA configuration; `python -m nemo_automodel.components.models.praxis.lora_export` merges the adapters into a checkpoint.
Prepare your environment	Follow the installation instructions.

Configuration

Setting	Configuration
Hardware	8 x H100 80 GB (one node)
Strategy	FSDP2 with expert parallelism; ep_size=8, tp_size=1, pp_size=1, cp_size=1
Nodes	1
Features	Checkpointing (consolidated checkpoint at every save), Activation checkpointing, Transformer Engine grouped experts with HybridEP token dispatch, fp32 optimizer state
Advanced	param_dtype=bfloat16, reduce_dtype=float32, dispatcher=hybridep

Results

The tutorial fine-tunes the single-arm checkpoint on the LIBERO-Spatial demonstrations and evaluates it with the standard LIBERO protocol, 500 paired episodes per checkpoint.

Checkpoint	LIBERO-Spatial success
`runwayml/praxis-1-droid`, zero-shot	0/500
full fine-tune, 600 steps x 48 windows	451/500 (90.2 %)
LoRA, 600 steps x 48 windows	457/500 (91.4 %)

Model Reference

Model Architecture

Property	Value
Task	Robot policy (video prediction and action generation)
Architecture	`PraxisPolicy`

Available Models

runwayml/praxis-1-droid: the single-arm checkpoint; the recipe reads its config.json for the cameras, the action space and the normalization.

The recipe supports the single-arm embodiment; the bimanual checkpoint has a different action space and camera layout.