Report: Runway releases “Praxis-1,” a world action model for robots

The Robot Report reported on November 12, 2026, that Runway released “Praxis-1,” a world action model for robots. According to the report, the model weights are available to developers under the Apache 2.0 license.

Runway reportedly released the weights and code for inference and fine-tuning under Apache 2.0, with the expectation that developers will run and adapt the model on their own hardware, including for commercial use. For developers considering adoption, practical questions include the terms of use and whether they can verify the reproducibility of adjustments for their robots and tasks.

The main point of this development is less the announcement of a model for robots than the way its weights and inference and adjustment code are being made available to developers under Apache 2.0. According to the report, Runway says it has released a package that lets developers run and adapt the model on their own hardware, including for commercial use. Rather than having the model provider manage the physical robots and usage environments end to end, this approach appears to give users more scope to bring in and adjust the model themselves. However, the existence of a public license and code alone does not establish that performance or operating conditions will be the same across different robots. In an article the previous day, we described Praxis-1 as a 27B-parameter model that predicts video and actions simultaneously, and noted its 46.2% success rate on RoboLab-120, as announced by Runway Runway releases “Praxis-1,” connecting video pretraining to robot control. Rather than reexamining the model’s specifications or evaluation results, this report shifts the focus to use on physical robots after release and to how the model is distributed so developers can adapt it to their own environments. Treating these separately helps avoid conflating benchmark results with implementation and adjustment for individual tasks. In robot learning, data collection, model training and evaluation on physical robots are distinct stages. Our coverage of a report on a standardized pipeline for video data also offers context for considering this separation Memories.ai reportedly builds a standardized pipeline for first-person video data for robots. Releasing weights and adjustment code provides an entry point for using the model, but does not by itself reveal what data was used for additional training or how evaluations can be aligned when tasks or robots differ. Points to check going forward include third-party replication results using the released code, the amount of data needed to adapt the model to each robot, and evaluations of continued operation on physical robots. These details cannot be established from the reported release information; please consult the original source for further details.