Mimic Robotics Announces New Class of Video-Action Models Based on Video Models

Mimic Robotics announced the release of mimic-video in a blog post. The company argues that video-model backbones are a more natural choice than VLM backbones for pretraining robot foundation models.

According to Mimic Robotics, mimic-video is a new approach that builds robot policies on pretrained video models. Rather than relying on image- and language-centered foundations, it uses representations designed for video generation, reconsidering how physical motion and coherent actions are handled.