The robotics startup’s latest AI framework specializes in rapid skill acquisition for industrial automation, allowing robots to adapt to unfamiliar factory workflows using minimal training. Technical details released for the V0.5 open-weight model show that it can acquire new industrial tasks from less than 2 hours of demonstrations. Trained on more than 1.7M demonstrations using the Wan2.2 video generator, V0.5 outperformed a rival model in environmental stability tests; when faced with randomized clutter and lighting on the RoboTwin 2.0 benchmark, V0.5’s success rate fell to 73.4%, while Physical Intelligence’s π0.5 dropped from 70.7% to 46.0%.
Verne is now scaling its training data 10x after finding that increasing pretraining data without task-specific fine-tuning increased the model’s benchmark success rate from 23% to 51%. To ensure reliability in industrial settings, the model runs inference directly on the robot hardware to remove the need for constant cloud connectivity or high-speed Wi-Fi. The company has published the weights and inference code for V0.5 on Huggingface.
