PlayJev incontra la robotica -PlayJev x Libero Robot control with Jev-like model
Can a small multimodal model—originally built to play video games from pixels—learn to control a robot?
In our new blog post, we explore how we adapted PlayJev, a 0.8-billion-parameter model, for closed-loop robotic control on LIBERO-Spatial. Given camera observations and a language instruction, the model predicts the robot’s actions in a single forward pass.
The result: 492 successful episodes out of 500, achieving a 98.4% success rate under the official LIBERO-Spatial evaluation protocol.
The article presents the model’s architecture, its discrete action representation, and real examples of the robot completing manipulation tasks.
