AI & Computer Science · Weekly AI Papers
La-JEPA: Compact Robot Policies with Latent World Modeling (Mateo Clémente et al.)
LISTEN TO THIS EPISODE

Mateo Clémente and colleagues introduce La-JEPA, a compact robot policy trained in the feature space of a pretrained V-JEPA 2.1 video encoder.
The model learns to predict future visual features alongside actions during training, then generates actions alone when controlling the robot.
The authors evaluate it on LIBERO-Plus, RoboTwin 2.0 and four real-robot tasks, reporting strong results with smaller policies than several generation-based alternatives.
Their design study compares JEPA features with video-VAE latents and examines which forms of predictive supervision improve control.
The paper also identifies corrupted instructions in the released LIBERO-Plus evaluator and re-evaluates baselines under a corrected protocol.
Read about the original source: La-JEPA: Compact Robot Policies with Latent World Modeling
Part of Marginalia: books, papers, and ideas for learning and debate.