OSCAR robotics world model debuts

By standardizing robot actions into 2D kinematic skeletons, OSCAR bypasses the hard problem of multi-embodiment translation, proving that cross-platform robotics models can be trained efficiently on modest hardware.

* Embodiment Agnosticism: Leveraging 2D skeleton joint maps abstracts away the complex differences in robot geometry, offering a unified representation for arms, humanoids, and hands.

* Democratized Compute: Achieving state-of-the-art world modeling on a single GH200 GPU shows that massive supercomputing clusters are not strictly required for meaningful robotic world models.

* Video-Based Simulation: By predicting future video frames based on actions, it functions as a visual simulator, helping debug and evaluate control policies before physical deployment.