LeRobot v0.6.0 drops world model policies

World models are transitioning from theoretical concepts to resource-efficient, practical deployment frameworks. By training VLA-JEPA with a latent video world model that is discarded at runtime, developers can bypass the compute-heavy inference bottlenecks typical of standard vision-language-action models. Furthermore, LeRobot lowers the barrier to entry by enabling fine-tuning on as few as 13 trajectories and integrates these advanced policies alongside low-cost hardware support to accelerate real-world deployment.