Here's my two cents as a mere paper reader; JEPA is really just a generalized encoder, so the JEPA created latents should be fed into a transformer trained wit

Here's my two cents as a mere paper reader; JEPA is really just a generalized encoder, so the JEPA created latents should be fed into a transformer trained with either user data or RL policy. Although the above might not work great either because you said that vertical position was not predicted well! Great article and I hope that you can carry on with the JEPA research