Clarity for AI buildersSponsored by AI Routing
availability

Runway releases Praxis-1 robot-control weights and code, with company-run test results

Runway releases Praxis-1 robot-control weights and code, with company-run test results

Runway has released Praxis-1's weights and code for robot control, according to The Decoder.

The release expands access beyond the selected partners that could use the model after its introduction in late September. Runway also published a research report with additional performance tests. The weights and source code carry an Apache 2.0 license, which permits commercial use, The Decoder reports.

Praxis-1 applies video-model training to robot control. It takes camera images, text instructions and information about a robot's position, predicts changes in the environment, and calculates corresponding movements. Incoming camera frames let it continually adjust its control. Runway calls this approach a World Action Model.

The underlying video model learned how objects move and interact from ordinary footage before further training on human-action videos and robot data. Runway, whose video models include Gen-4.5, is using that foundation to extend its video research into physical control. According to The Decoder, Standard Bots, Noble Machines and Ultra have been testing Praxis-1 on systems including a humanoid, an industrial arm and a dual-arm warehouse robot.

Video pretraining does not eliminate robot-specific training

The potential benefit is a reduced need for costly robot-specific datasets, but the supplied report does not establish that Praxis-1 needs less robot data overall than comparable models without general video pretraining.

For adaptation to a Franka robot arm, Runway used about 277 hours of recordings from the public DROID research dataset. A dual-arm robot required between 1,400 and 2,600 hours of training material. DROID recordings and data from other robot systems had also been used in earlier training stages.

That distinction matters when assessing the approach: the reported adaptation hours are not a complete accounting of the robot data used to develop the model. The release makes a pretrained foundation available, but it does not demonstrate that developers can dispense with collecting or obtaining data for their target hardware.

Benchmark gains come with verification limits

On Nvidia's RoboLab-120 benchmark, which covers 120 simulated tasks involving actions such as sorting, grasping and stacking, Praxis-1 achieved a 46.2 percent success rate. FLUX 3 Action, another World Action Model, scored 42.9 percent under the same test conditions, according to The Decoder.

Praxis-1 led the World Action Model category but placed fifth overall, behind two AI agents and two robot models with different architectures. Its category lead therefore should not be read as an overall lead across robot-control approaches. Runway conducted the tests itself, and the results have not been independently verified.

In real-world Franka tests following DROID fine-tuning, Praxis-1 completed 88 of 100 attempts, compared with 80 for FLUX 3 Action. Each model attempted the same ten tasks ten times, including placing a banana on a plate and closing a drawer.

Runway also tested adaptation to a mobile robot platform with a gripper arm previously used by Google to develop RT-1. That platform's recordings were absent from Praxis-1's earlier training. After fine-tuning on about 95 hours of robot data, Praxis-1 achieved an average 80 percent success rate in simulation across four task types. Google's original RT-1 scored 52.4 percent on the same benchmark.

These results address different questions: performance on a simulated task suite, execution on a physical arm, and adaptation to an unfamiliar robot platform. Taken together, they provide evidence of capability in the tested settings, not a single measure of reliability across robots or deployment environments.

Hardware and planning remain constraints

Developers can obtain the weights through Hugging Face after registering and accepting access terms; the code is on GitHub. The Decoder reports that pre-adapted versions are available alongside a base model that needs additional training for new robot systems.

The standard version requires a GPU with 80 GB of memory. Runway estimates that a compressed version should fit on an RTX 5090 with 32 GB of VRAM, but the GitHub documentation says actual peak memory use has not yet been measured. That estimate is therefore not a verified hardware requirement for deployment.

According to Runway's research report as described by The Decoder, Praxis-1 can confuse similar-looking objects and struggle with vague instructions. Longer tasks require explicit individual steps because it cannot yet reliably plan complex action sequences. Its current role is consequently closer to executing specified movements than independently organizing a multi-step job.

Runway suggests that general-purpose AI models could eventually supply that higher-level planning, but the report presents this as a future possibility rather than a demonstrated capability.