SkyJEPA

Summary

SkyJEPA is a JEPA-style latent dynamics model for real-time quadrotor control. It predicts future latent dynamics from state and motor-control histories, maps frozen latent rollouts back to physically meaningful state with a physics-inspired prober, and uses the result inside an MPPI controller for zero-shot sim-to-real outdoor flight.

Role In The Wiki

SkyJEPA is the quadrotor-control counterpart to the wiki’s JEPA/world-model line. Compared with VLA-JEPA, it is less about generalist vision-language-action policy pretraining and more about an explicit state-action dynamics model used for candidate control-input rollout. Compared with LeWorldModel, it adds a real-time aerial robotics deployment and a structured prober for metric-state recovery.

Evidence

Official Artifacts

Artifact caveat: the official GitHub repository currently says source code, pretrained models, dataset instructions, and reproduction scripts will be released soon.

Relation To Foundation TSFM Agenda

Use the source-level agenda mapping in skyjepa-2026 rather than duplicating verdict rows here.

At the entity level, SkyJEPA is a named physical-robotics world model whose transferable lesson is the interface pattern:

state history + control-input history + candidate control sequence
  -> latent rollout
  -> structured physical-state rollout
  -> receding-horizon control choice

For the foundation time-series agenda, this is strong physical-robotics analogy evidence for action-conditioned latent dynamics and structured readout, not direct evidence for digital-world telemetry or general multivariate time-series foundation models.