Crate jepa_world 

jepa-world

World model primitives for JEPA-based planning and control.

This crate bridges the gap between learned representations (from jepa_core / jepa-vision) and decision-making. It provides:

• Action-conditioned dynamics — predict next-state representations given an action, enabling model-based rollout (RFC-009).
• Hierarchical JEPA (H-JEPA) — stack multiple JEPA levels at different temporal/spatial scales (RFC-010).
• Short-term memory — bounded FIFO buffer of recent states for multi-step planning context.
• Planning — RandomShootingPlanner (CEM) evaluates candidate action sequences against a learned world model and cost function.

                 ┌───────────────┐
observation ────►│   Encoder     │──► s_t
                 └───────────────┘     │
                                       ▼
                 ┌───────────────┐   ┌──────────────────────┐
         a_t ───►│  Dynamics     │──►│  s_{t+1} ... s_{t+H} │──► CostFunction ──► plan
                 │  (World Model)│   └──────────────────────┘
                 └───────────────┘

Modules

Module	Contents
action	Action wrapper, ActionConditionedPredictor trait
planner	WorldModel, RandomShootingPlanner (CEM), CostFunction trait, L2Cost
hierarchy	HierarchicalJepa, JepaLevel — multi-scale H-JEPA
memory	ShortTermMemory — bounded ring buffer of recent states

Re-exports

pub use action::Action;

pub use action::ActionConditionedPredictor;

pub use hierarchy::HierarchicalJepa;

pub use hierarchy::HierarchyError;

pub use hierarchy::JepaLevel;

pub use memory::MemoryError;

pub use memory::ShortTermMemory;

pub use planner::CostFunction;

pub use planner::L2Cost;

pub use planner::PlanResult;

pub use planner::PlanningError;

pub use planner::RandomShootingConfig;

pub use planner::RandomShootingPlanner;

pub use planner::WorldModel;

Modules

action

Action types and action-conditioned prediction for world models.

hierarchy

Hierarchical JEPA (H-JEPA) for multi-scale prediction.

memory

Short-term memory module for world models.

planner

World model planning via rollout and plan evaluation.