Runway releases Praxis-1 weights for the SO-101 tabletop robot arm

Runway said on Nov. 13 that Praxis-1 can run on the SO-101, a tabletop robot arm that researchers and hobbyists can assemble. The video AI company published two sets of weights trained for the SO-100 and SO-101 arms, along with code for running them. Runway says the version prepared for NVFP4, a lower-precision format, can run on a single GeForce RTX 5090.

That gives teams with an SO-101 a version of Praxis-1 already trained to control their arm. The Nov. 11 release included weights for a single-arm research setup and a dual-arm setup, plus a base model that teams could adapt to new equipment. The Nov. 11 GitHub update listed SO-101 support as upcoming. A later change added weights and code that target the joints and gripper of the SO-100 and SO-101.

Training on recordings shared by robot owners

The SO-101 is an open-design robot arm for experiments such as picking up and moving objects on a workbench. Hugging Face’s LeRobot project publishes a parts list and 3D printing resources. During demonstrations, one arm can be used to guide the other. The new weights control the follower arm, which receives and carries out the commands.

Runway gathered about 1,260 hours of recordings in 3,956 public LeRobot datasets from 1,783 contributors, according to the technical report. It used about 1,070 hours from 3,814 datasets to train the SO-101 version. The team set aside 102 complete datasets for validation so it could check the model on recordings kept out of that training.

The researchers also found that labels on the recordings did a poor job of identifying where a camera was mounted. They examined what moved in the footage and what the scene looked like to distinguish cameras fixed to the workbench from cameras that moved with the wrist. The training included examples with only a view of the work area, so the model can take that view alone or use it alongside a wrist camera.

The arm receives targets for five joints and a gripper

Praxis-1 receives a task instruction, the camera view and the position of the arm. It calculates targets for the five joints and the gripper. The report describes 15 sets of joint and gripper targets per second, generated in groups covering about 1.6 seconds of movement. That rate describes the spacing of the targets. How long the arm takes to complete a task depends on the object and the movement involved.

Runway shows the arm folding a towel and wiping a plate in its demonstrations. The SO-101 section of the report gives no success rate across repeated trials. The NVIDIA simulation result that Runway cited for its earlier Praxis-1 release used weights for a different arm.

Both versions still require a 54.4 GB download

Runway trained the NVFP4 version to account for the lower precision used in part of the model at runtime. Both versions are distributed as 54.4 GB of 16-bit weights. The lower-precision version packs part of those weights when they are loaded. Runway’s repository puts the resulting memory for the control model’s weights at 17.5 GB, down from 54.4 GB before packing.

The separate text and image-processing weights add 4.6 GB and 1.4 GB, respectively. That brings the estimated total for weights to about 23.5 GB before memory for intermediate computation. Runway says its SO-101 version runs on one RTX 5090, and the repository’s memory explanation leaves peak usage unmeasured. The robot and its camera still have to supply the observations that drive each round of movements.

Both Hugging Face pages list the weights under Apache 2.0 and require users to sign in and consent to the access terms, which include sharing account information. The text and image components are distributed separately under their own licenses.