Action Space

I believe actions should be generated by models that deeply learn the action modality itself, rather than relying on priors from video or vision-language models. We need these action models, along with benchmarks to evaluate them. With strong action models in place, the next question is how to use t…