Runway's Praxis-1 Beats Every Open Robot Model at 46% Success Rate

Runway has released Praxis-1, a world action model that converts predicted video into robot commands. The model uses general internet video to learn motion and object interaction, then adds robot-specific training to map that knowledge to physical actions. Its weights, inference code, and full and low-rank adaptation fine-tuning code are available under Apache 2.0, including for commercial use.

The release is available through GitHub, Hugging Face with native LeRobot support, and vLLM-Omni. Runway also offers hosted inference through Reactor. Three checkpoints are included: a base mid-trained model, a version fine-tuned on DROID, and one fine-tuned for the bimanual YAM robot.

From future frames to motor commands

Praxis-1 combines three components: a video backbone that predicts future camera frames, an action expert that extracts the motion needed to produce those frames, and per-robot projections that map different pose and action formats into a shared representation. Each policy call receives an instruction, recent camera frames, and the robot’s current pose. It returns predicted video and 1.6 seconds of control commands, represented as 24 action rows at 15 Hz.

The video backbone is a diffusion transformer derived from Runway’s video foundation model. Seventeen of its 20 feed-forward layers use a mixture-of-experts design. Each layer contains one shared expert and 128 routed experts; every token passes through the shared expert and eight selected routed experts. This routing expands model capacity while keeping per-token computation near that of a dense layer. The backbone contains 26.9 billion parameters, with about 2.8 billion active for each token.

The action expert is a 16-layer dense transformer with about 314 million parameters. It cross-attends to representations from eight depths of the video backbone, allowing it to recover actions from both low-level motion and higher-level scene features. Action-loss gradients also update the backbone during training.

Four passes from video to control

1. Video pretraining. The backbone begins as Runway’s video foundation model, trained on general video across 12 generation tasks, including text-to-video and reference-conditioned generation. Full-sequence attention allows information to move across the complete training clip.
2. Embodied-video mid-training. The model switches to chunk-causal attention, which organizes attention around temporal blocks for sequential prediction. The data mix is roughly 58% egocentric human video, 20% general video, 18% real-robot video, and less than 1% gameplay, sampled at 15 frames per second.
3. Joint video and action mid-training. A newly initialized action expert receives action targets for robot examples. Training uses an equal-weight combination of video and action velocity losses, which teach each branch to recover its target from noise. Noise levels are sampled independently, and about 30% of tasks remain video-only.
4. Robot-specific fine-tuning. The complete model is adapted to one embodiment. For the DROID Franka checkpoint, Runway trained for 50,000 steps on 128 H100 GPUs. Only the selected robot’s input and output projections begin with fresh weights.

Runway’s scaling study found that additional mid-training compute improved later fine-tuning on robots both present and absent from the training mix. RT-1 was excluded from mid-training, yet its SimplerEnv success rate increased from 56.3% with the earliest checkpoint to 80.0% with the latest.

Strong open-weight results, with caveats

On NVIDIA’s RoboLab-120 leaderboard, the DROID-tuned Praxis-1 completed 555 of 1,200 episodes, a 46.2% success rate. Runway reports that this is the highest result among listed world action models and open-source robot foundation models.

Four leaderboard entries score higher overall, although the leading systems use a different inference setup. AMP1, at 86.0%, and SenseNova-RoboRSI, at 57.1%, query frontier models during execution instead of running a trained policy alone. SenseNova-RoboRSI’s authors report about 55 minutes of computation per episode for 8 to 76 seconds of commanded motion. Praxis-1 generates each 1.6-second action chunk in about 0.55 to 0.6 seconds on one H100.

Runway also tested Praxis-1 against GPT-6 Astra on the same tasks using the same bimanual YAM robot. Its reported averages show Astra taking 15.6 times longer to complete a task and costing 51 times more per successful completion. In a “remove the pen cap” episode, Praxis-1 finished in 13 seconds; Astra took 7 minutes and 54 seconds.

A new robot from 29 minutes of demonstrations

For Noble Machines’ robot, Runway fine-tuned Praxis-1 on 154 teleoperated episodes totaling about 29 minutes. Training took 4,000 steps, or roughly 3.6 hours on 16 GPUs. The resulting policy succeeded in all 15 simulated trials across two layouts represented in the training data. On an unseen warehouse task, it placed 25 of 30 planks across 10 episodes.

RT-1 provides a separate test of adaptation because that robot never appeared during mid-training. After fine-tuning on RT-1 data, Praxis-1 averaged 80.0% across four SimplerEnv tasks. RT-1 itself scored 52.4%, and FLUX 3 Action reached 39.7% under the same fine-tuning recipe.

One 80 GB GPU, with a smaller option

The complete model contains 27.2 billion parameters and occupies 54.4 GB in bfloat16 format, allowing it to run on one 80 GB GPU. Runway also quantized the routed experts with Nvidia’s four-bit NVFP4 format, reducing policy storage to 17.5 GB. In a separate RoboLab-120 comparison, the quantized model scored 46.5%, versus 47.8% for the bfloat16 version; Runway reports that the difference was not statistically significant.

Supported Nvidia hardware ranges from H100 and B200 data-center GPUs to RTX 5090 and RTX PRO 6000 workstation cards. Jetson Thor can run the policy aboard a robot. The release also includes a policy server, a MuJoCo simulation environment, and code for full fine-tuning and LoRA, a parameter-efficient method that updates small adapter matrices.

Failures begin in imagined futures

Runway’s technical report identifies three recurring failure modes:

• Look-alike objects. The most common real-robot error is selecting an object that resembles the requested target. The incorrect choice usually appears in the predicted video before the arm moves, locating the error in the model’s imagined plan.
• Vague instructions. Performance improves when prompts specify attributes such as color or size. Instructions without distinguishing details leave several plausible targets.
• Long-horizon tasks. The policy can execute multi-step tasks when each step receives a separate instruction. Following one prompt across several stages remains difficult, and Runway lists integration with a large language model planner on its roadmap.

An intervention in the report helps isolate the main source of error. Researchers supplied the true future video, added noise matching the sampler’s operating level, and generated only the actions. Action error fell to roughly one-fifth of the normal deployment error. The result attributes most remaining mistakes to future-video prediction and suggests that better video generation could improve control without redesigning the action expert.

The bet on video-scale robotics

Many generalist robot policies use vision-language-action architectures built on vision-language models, inheriting semantic knowledge from image and text pretraining. Praxis-1 uses temporal prediction as its central prior, drawing on video to learn how hands, tools, and objects move through a scene. Its reported scaling results suggest that broad video pretraining can reduce the amount of robot-specific demonstration data needed for a new embodiment.

Apache 2.0 weights, commercial-use terms, multiple checkpoints, and complete fine-tuning code make the architecture available for independent deployment and evaluation. Its broader value will depend on whether the reported sample efficiency and latency persist across more robots, tasks, and uncontrolled environments.