From video to robot motion: Praxis-1 is released with open weights

Released with open weights under Apache 2.0, Praxis-1 predicts video and actions. Its results distinguish simulation from tests on physical robots.

A base model to adapt, one version for a Franka arm, and another for a dual-arm YAM robot: the Praxis-1 release provides access to Runway’s robot control models. The code includes inference, a policy server, a MuJoCo simulation, and tools for full fine-tuning or LoRA adaptation. The base model, which lacks robot-specific interfaces, must be adapted before it can control a robot.

The code repository and weights are distributed under Apache 2.0. Third-party components, including the Qwen3.5-2B text encoder and Wan2.2 VAE, remain subject to their respective licenses. Standard inference requires a GPU with at least 80 GB of memory; the policy weights alone occupy approximately 54.4 GB in BF16. Predicting the next scene and the movements that produce it

In the technical report, the World Action Model combines a video model with an action expert. Given an instruction, camera images, and the robot’s pose, it jointly predicts the visual sequence and movement commands. Each standard chunk spans 1.6 seconds, with 24 commands at 15 Hz. The robot’s actual observations then feed into the next call.

The video backbone has 26.9 billion parameters, with approximately 2.8 billion active per token through a mixture-of-experts architecture. The action module adds approximately 314 million parameters. Training starts with general video, then introduces manipulation-related videos and action data before adaptation to a particular robot. The release therefore does not describe learning robot control solely from ordinary videos. First among World Action Models

The official NVIDIA RoboLab-120 leaderboard lists 555 successful episodes out of 1,200 for Praxis-1, with a reported score of 46.2%. The benchmark includes 120 manipulation tasks in simulation, with 10 trials per task for this entry. The model leads the other World Action Models on the board, including FLUX 3 Action at 42.9%.

This first-place ranking applies to a category. On the overall leaderboard reviewed, four systems score higher, placing Praxis-1 fifth. The top two, AMP1 at 86% and SenseNova-RoboRSI at 57.1%, are agents driven by general-purpose models. The “number one” claim should therefore not be extended to all evaluated approaches, and the benchmark score remains a simulation measurement. Adapting the model to a robot still requires training

For Noble Machines’ robot, the team reports 15 successful simulated trials out of 15, split between two layouts present in the training data. Teleoperated demonstrations total 29 minutes, with 26 used for training. Fine-tuning takes approximately 3.6 hours on 16 GPUs. On a harder task absent from those demonstrations, six out of 10 episodes succeed completely. The report specifies that this test has not yet been run on the physical robot.

The experiment involving Google’s RT-1 robot uses a different scale: approximately 95 hours of data and 28 hours of fine-tuning on 32 H100s. The result reaches 80% in SimplerEnv across four tasks and 864 episodes, compared with 52.4% for the reference RT-1 policy. The robot was absent from pretraining and mid-training but present in the adaptation data.

Physical demonstrations accompany these tests on Franka, YAM, and the SO-101 arm. Acknowledged difficulties include confusion between similar-looking objects, vague instructions, and long action sequences requested in a single prompt. For the latter, providing a separate instruction for each step remains more effective, according to the team. A speed comparison limited to three tasks

To measure execution time and cost, the company compares its policy with GPT-6 Astra on the same YAM station and three tasks: removing a pen cap, putting a sponge in a bag, and placing a die in it. The internal results report a 15.6-fold speedup and a 51-fold reduction in cost per completed task for Praxis-1. This protocol does not measure general superiority over language models.

Average time covers successful episodes. Cost per success also includes failed attempts, using an H100 priced at $3.95 per hour for Praxis-1 and Astra’s API bill. Training, demonstration