Verne Robotics releases V0.5, an open-weight world action model aimed at industrial robots

Verne Robotics, a San Francisco company founded in 2025, has published V0.5, which it describes as an open-weight world action model meant to speed up the deployment of industrial robots. The model was announced on November 4, according to AI Post Hub, and the weights are hosted under the company's Hugging Face organisation, with a GitHub repository linked from the announcement.

How the model works

A world action model learns from video how scenes change and uses that knowledge to choose actions. V0.5 combines two transformers, one for video and one for actions, trained together with a flow-matching objective. It starts from the Wan2.2 video model with an expanded architecture. At run time it takes camera images, memory, the robot's state and a language instruction, then outputs chunks of end-effector and gripper commands at 30 Hz without generating future video frames. In Verne's stack, an off-the-shelf vision-language model handles high-level planning, V0.5 acts as the policy, and a custom controller runs at 500 Hz.

Pretraining used more than 1.7 million demonstrations, mostly teleoperation and handheld-gripper recordings from vendors, open datasets and Verne's own collection. The company says that, without fine-tuning, success on RoboTwin 2.0 rose from 23% to 51% as it went from 29% to 100% of its pretraining data, and that it is now scaling data tenfold.

Reported results

On the LIBERO simulation suites, Verne reports a 99.3% average, compared with 96.85% for Physical Intelligence's π0.5 and 99.4% for ABot-M0.5. On RoboTwin 2.0, a set of 50 two-arm tasks, V0.5 scores 81.0% in clean scenes and 73.4% in randomised ones after fine-tuning on 2,500 demonstrations; π0.5 scores 70.7% and 46.0%. Verne says V0.5 trails π0.5 slightly on tasks where the latter already succeeds most of the time. The company calls V0.5 the strongest open-weight model from a US company on both benchmarks as of November 1.

Limits and open questions

The figures come from Verne and have not been independently reproduced. Real-world tasks such as bin picking, DIN rail assembly and pizza box folding are shown in videos, without success rates. The announcement does not state the licence or the parameter count, and it does not say whether the GitHub repository includes training code. Post-training is offered through a hosted platform that requires sign-up, alongside bespoke deployment services for partners.