Kepler — world encoder

A self-supervised world encoder for robots. Vision, proprioception, and force/torque are fused into a single 256-d latent by 8 learned queries (Perceiver-style bottleneck), trained JEPA-style: hide one sense, predict its embedding from the others. Trained on RH20T (~54M frames, 4 embodiments). One encoder transfers across robot bodies, and the vision-only latent linearly reads out motor / end-effector / force state — signals raw pixels alone don't carry ("sensory entanglement").

• Paper: https://arxiv.org/abs/2607.13522
• Blog: https://menlo.ai/research/kepler-v01
• Docs: all research docs — plan, experiment journals, live run docs (RH20T temporal, RoboCasa N1), results, paper source — on the docs branch

Code layout

• world_tokenizer/ — encoder, JEPA pretraining, probes, predictors
• preprocessing/ + run_*.sh — data prep and run entrypoints
• metrics/, visualizer/, scripts/, splits/