runwayml/praxis-1-so101

Praxis-1 SO-101

Praxis-1 policy fine-tuned for the SO-100 / SO-101 follower arm (LeRobot), in joint space.

Model	27.2B-parameter MoE video-diffusion policy: chunk-causal trunk + 0.3B action expert
Frozen dependencies (not in this repo)	Wan2.2 VAE, Qwen/Qwen3.5-2B text encoder
Embodiment	so101: one arm, cameras exterior + wrist (one or two views)
Canvas	288x416 tiles side by side (288x832 for two views); a 640x480 camera gives a 288x416 tile
Control	15 Hz, 24-step action chunks
Action space	6-D [q0..q4, g]: the 5 arm joints in radians, relative to the chunk-start joints, and the absolute gripper closure in 0..1 (1 = closed)
Pose frame	joints (joint angles, no end-effector pose)
Weights	bf16, 54.4 GB, 17 safetensors shards (model.safetensors.index.json)

Normalization statistics and the action codec are in config.json; no dataset statistics are needed at inference time.

The rig reads LeRobot SOFollower units with use_degrees=True: the joints in degrees and gripper.pos in the follower's 0..100 scale. Prompts go to the text encoder verbatim; send a short imperative task sentence, for example "Pick up the red cube and put it in the box."

Files

• config.json: praxis model config (model_type: praxis, format_version: 2)
• model-000NN-of-00017.safetensors + model.safetensors.index.json: sharded weights
• SHA256SUMS: checksums for every file above

Usage

from huggingface_hub import snapshot_download

path = snapshot_download("runwayml/praxis-1-so101")

Load path with runwayml/praxis-1-oss and serve it with configs/serve_so101.yaml (see "SO-101 rig" in its README). Inference needs one 80 GB GPU.

Related

• runwayml/praxis-1-base: base model, no action heads; the start of a fine-tune for a new rig
• runwayml/praxis-1-so101-nvfp4: praxis-1-so101 with NVFP4 quantization-aware training
• runwayml/praxis-1-droid: single-arm DROID policy
• runwayml/praxis-1-yam: bimanual YAM policy
• runwayml/praxis-1-oss: reference code (inference, serving, fine-tuning)

License

Apache License 2.0. The frozen dependencies (Wan2.2 VAE, Qwen/Qwen3.5-2B) are distributed separately under their own licenses.