Runway releases Praxis-1 robot policies under Apache 2.0

Runway has released Praxis-1’s model weights and inference and fine-tuning code under Apache 2.0, opening its robot-control model to developers beyond the partners testing it since September.

The 11 November announcement makes the change in access explicit. Runway’s September introduction offered selected partners early access and promised a public release in the following months. The public repositories now include downloadable weights, rather than just that promise.

Praxis-1 predicts a short sequence of video and robot actions together. Given a task instruction, camera views and the arm’s current state, it generates the movements for the robot to execute. Its released DROID policy has 27.2 billion parameters and produces action commands at 15 Hz.

What developers can download

Runway’s Hugging Face collection contains policies for a single-arm DROID rig and a two-arm YAM setup, plus a base checkpoint for adaptation to other robots. The base has no action heads: developers must initialise and train those components for their own rig.

The DROID files comprise 17 safetensors shards totalling 54.4 GB, with configuration data and checksums. Its licence file and the code repository use Apache 2.0. The separately downloaded Wan2.2 video autoencoder and Qwen3.5 text encoder retain their own licences.

The code includes serving, simulation and full or LoRA fine-tuning. LeRobot-format checkpoints are also listed in the collection.

That gives developers a different starting point from NVIDIA’s recently announced GR00T 2, whose code was promised for December. The distinction is release readiness, not a comparison of how reliably the two models control a robot.

Desktop GPU support is still experimental

The reference installation calls for Python 3.12 and a CUDA GPU with at least 80 GB of memory. The DROID model card specifies the same memory requirement for inference and its single-GPU LoRA adaptation example.

Runway’s launch post also lists the RTX 5090 among supported hardware. The repository qualifies that claim: its lower-memory NVFP4 path for RTX 5090 and RTX PRO 6000 Blackwell is experimental. It has not yet been tested in closed-loop operation on those SM120 GPUs, and the documentation says speed and peak memory remain unmeasured.

Runway estimates about 23.5 GB for the quantised policy and its frozen dependencies before activations. It says a 32 GB RTX 5090 should fit, while a 24 GB card does not. Those are documented estimates, rather than measurements from Sulat.

Availability does not settle reliability

The technical report identifies confusion between similar-looking objects, difficulty with vague instructions and trouble completing long sequences from one prompt. Multi-step tasks work better when each step receives its own instruction.

The release makes those behaviours inspectable and adaptable through public code and weights. It does not establish dependable operation on every robot or in an untested physical environment.