D-JEPA
A shared start. Different actions.RECORDED EXECUTION

00:00 / 01:20

LEARNING FROM EXECUTED OUTCOMES

Teach the relations that determine the action.

Candidate futures→Executed outcomes→Relational alignment

Successful candidates receive decision mass. Local success–failure pairs sharpen the action boundary; the correction penalty keeps the update localized.

L = Lsuccess + λlocalLlocal + λtrust‖δ‖²

Outcomes supervise training. At selection time, the operator consumes predictive evidence.

CROSS-TASK VALIDATION

Open original video ↗