Tag
This paper investigates whether action-conditioned JEPA world models can perform latent planning with point cloud observations instead of images, finding that they succeed and discussing implications for geometric perception in robotics.