| name | neural-fields-world-models |
| description | Neural Fields as World Models methodology — isomorphic world models that preserve sensory topology for physics prediction as geometric propagation rather than abstract state transition. Motor-gated neural fields with local lateral connectivity and action-conditional prediction within spatial maps. Use for: world model architectures, sensory cortex modeling, offline task learning, action-conditional prediction, spatial prediction, embodied AI, neural field implementations. Activation: neural field, world model, isomorphic, spatial topology, motor-gated, action-conditional, offline learning, embodied cognition, sensory preservation. |
| license | Complete terms in LICENSE.txt |
| metadata | {"arxiv_id":"2602.18690","published":"2026-06-01","authors":"Joshua Nunley","tags":["neural-fields","world-models","spatial-prediction","embodied-cognition","offline-learning","motor-gated","isomorphic"]} |
Neural Fields as World Models
Isomorphic world models that preserve sensory topology, enabling physics prediction as geometric propagation rather than abstract state transition.
Core Concept
Traditional world models compress visual input into latent vectors, discarding spatial structure that characterizes sensory cortex. This paper proposes isomorphic world models — architectures that preserve sensory topology so prediction becomes geometric propagation.
Key insight: Physical prediction, offline task learning, and body-linked representation share a common computational substrate: action-conditional prediction within a spatial map.
Methodology
Motor-Gated Neural Fields
- Architecture: Activity evolves through local lateral connectivity
- Motor modulation: Motor commands multiplicatively modulate specific channels
- Spatial preservation: Sensory topology maintained throughout processing
Three Experiments
-
Ballistic prediction without teleporting
- Learns motion trajectories without instantaneous jumps
- Spatial continuity preserved
-
Offline task learning
- Catching policy improved offline
- Task error propagated through frozen learned world model
-
Body-selective motor channels
- Emerges without body labels
- Self-organized body representation
Key Features
Isomorphic Architecture
- Topological preservation: Spatial structure maintained unlike latent vector compression
- Geometric propagation: Physics prediction as spatial evolution
- Action-conditional: Motor commands gate field evolution
Motor Channel Organization
- Multiplicative modulation: Motor commands scale specific channels
- Body-linked emergence: Selective channels develop without explicit supervision
- Local connectivity: Lateral interactions preserve spatial relationships
Applications
Offline Learning
- Task improvement: Policy refinement without environment interaction
- Error backpropagation: Through frozen world model
- Mental rehearsal: Simulated practice through field dynamics
Embodied AI
- Spatial prediction: Motion trajectories in physical space
- Body representation: Emergent body-selective channels
- Action-conditional: Motor gating for goal-directed behavior
Sensory Cortex Modeling
- Topological structure: Preserves cortical organization principles
- Local interactions: Lateral connectivity mimics cortical circuits
- Prediction substrate: Shared foundation for multiple cognitive functions
Implementation Patterns
Neural Field Architecture
Input → Spatial Field → Local Lateral Connections → Motor-Gated Channels → Output
↑ |
|______________________________________________________|
Key components:
- Spatial field maintains topological structure
- Lateral connections enable local propagation
- Motor gating multiplicatively modulates specific channels
- Feedback loop for continuous prediction
Motor-Gating Mechanism
- Channel selection: Motor commands activate specific field regions
- Multiplicative scaling: Field values scaled by motor signals
- Selective propagation: Enhanced regions dominate evolution
Advantages Over Latent Vector World Models
| Feature | Latent Vectors | Isomorphic Fields |
|---|
| Spatial structure | Discarded | Preserved |
| Prediction type | Abstract state transition | Geometric propagation |
| Body representation | Explicit labels | Self-organized |
| Offline learning | Limited | Effective backpropagation |
| Teleporting artifacts | Common | Avoided |
Experimental Validation
Ballistic Prediction
- Success: Learns smooth trajectories without teleporting
- Baseline comparison: Standard world models show instantaneous jumps
Offline Task Improvement
- Performance: Catching policy enhanced through frozen model propagation
- Learning efficiency: Offline practice effective without environment
Body Channel Emergence
- Discovery: Body-selective channels emerge unsupervised
- Significance: Demonstrates self-organized body representation
Relation to Neuroscience
Cortical Principles
- Sensory topology preservation: Mirrors cortical spatial organization
- Local lateral connectivity: Matches cortical circuit structure
- Motor modulation: Similar to motor cortex gating mechanisms
Behavioral Analogies
- Mental practice: Offline rehearsal for skill improvement
- Dreaming: Action-conditional prediction in sleep
- Motor imagery: Spatial prediction without execution
Pitfalls
Computational Cost
- Spatial resolution: High-dimensional fields require substantial memory
- Lateral connections: Dense connectivity increases computation
- Motor gating: Channel modulation overhead
Training Challenges
- Spatial continuity: Requires careful regularization
- Body emergence: Needs sufficient motor variety
- Offline propagation: Frozen model must be stable
Implementation Issues
- Field initialization: Poor initialization disrupts topology
- Channel balance: Motor gating must avoid channel collapse
- Teleporting avoidance: Spatial propagation needs tuning
Activation Keywords
- neural field world model
- isomorphic architecture
- spatial prediction
- motor-gated neural field
- action-conditional prediction
- offline task learning
- body representation emergence
- sensory topology preservation
- geometric propagation
- embodied world model
Related Skills
- predictive-coding: Hierarchical prediction frameworks
- worldkv-world-memory: World models for memory
- hippocampal-entorhinal-world-model: Brain-inspired world models
- energy-based-neurocomputation: Energy-based prediction
- neuromechanical-locomotion-dynamics: Motor dynamics modeling
References
- arXiv:2602.18690 - Neural Fields as World Models (Nunley, 2026)
- Neural field theory literature
- World model architectures
- Embodied cognition research
- Cortical spatial organization studies