Video-Action Embodied Model Trained From Scratch

omarsar0 · x · 2026-07-11

The post argues that models trained purely to generate "good-looking video" primarily learn appearance, not how actions affect the world; directly adapting a generative model into a control model loses the priors accumulated during pre-training. To solve this, @robbyantbrain trained from scratch, natively building a complete video-action stack for embodied control. A reply mentions that world states and latent actions are mapped into the same semantic latent space, with actions learned self-supervisedly from unlabeled video, allowing even web videos to provide control signals.

Related event: LingBot-VA/VLA 2.0 Released: Native Embodied Foundation Model(24 posts)→

Original post →

More from Embodied

Embodied channel →