Machine Learning pathway increased my salary by 40%
Topic 9 of 12
Advanced
9-12 hours
Digital Twins & Spatial Computing
World Models: Environment Dynamics, Prediction and Planning for Physical AI
Master world model architecture, learned environment dynamics, prediction, planning, and the relationship to digital twins, LLMs, and Physical AI. Comprehensive guide covering model-based RL, generative world models, and embodied AI.
World ModelsModel-Based RLLatent DynamicsGenerative World ModelsVideo Prediction3D World ModelsEmbodied AIVLA ModelsFoundation ModelsPhysical AI
Executive Summary
World models represent one of the most significant emerging research directions in AI. They are models capable of understanding inputs from text, images, video, and other modalities to perceive the physical world and predict its future spatiotemporal dynamics. World models aim to learn or represent environment dynamics in a way that supports prediction, planning, and interaction. This is distinct from LLMs (which model language), VLMs (which model vision-language), and digital twins (which represent specific physical assets). The emerging research direction frames digital twins as potentially evolving from passive representations toward AI-enabled systems capable of prediction, intervention, and eventually more autonomous management. Proponents believe world models will overcome LLM limitations and realize AI's promise for robotics. However, this is an emerging research direction, not a universal production architecture.
Definition
A world model is a learned representation of an environment, its dynamics, and possible future states that can support prediction, planning, or action. It is distinct from a digital twin (which represents a specific physical asset with real-time synchronization), an LLM (which models language), a VLM (which models vision-language), a foundation model (which is a general-purpose large-scale model), and a simulator (which explicitly models physics). World models can be learned rather than explicitly modeled, enabling AI systems to develop internal representations of how environments work and use these representations for prediction, planning, and action.
Why It Matters
The future of AI is moving from language models to multimodal models to spatial models to world models to embodied/physical AI. World models represent the frontier where AI systems develop understanding of physical environments, their dynamics, and the consequences of actions. This is critical for robotics, autonomous vehicles, and any AI system that must reason about and act within physical environments. Recent research explicitly examines the transition from traditional digital twins toward AI-enabled twins and world models. The World Economic Forum has highlighted technological convergence as a transformative force. Enterprise world models represent an emerging research direction, distinct from general-purpose LLMs and digital twins. Enterprises face the dual challenge of ensuring knowledge retrieval accuracy and alignment with legal compliance and brand identity.
Evolution of the Technology
Language Models (2020s): LLMs model language and text. GPT, Claude, LLaMA. Vision-Language Models (2022+): VLMs combine vision and language. GPT-4V, Gemini. Spatial Models (2024+): Models that understand 3D space and spatial relationships. World Models (2025+): Models that learn environment dynamics and can predict future states. NVIDIA Cosmos, research models. Embodied AI / Physical AI (2026+): AI systems that perceive, reason, and act in physical environments using world models. Autonomous Physical AI (Emerging): AI agents that use world models for autonomous planning and action in physical environments.
Learning Objectives
- Understand what a world model is and how it differs from LLMs, VLMs, digital twins, and simulators
- Master world model architecture: perception, state representation, dynamics, prediction, planning
- Understand model-based reinforcement learning and latent dynamics models
- Explore generative world models, video prediction, and 3D world models
- Understand the relationship between world models and embodied AI/robotics
- Distinguish world models from digital twins and understand their complementary roles
- Understand world model validation: prediction error, long-horizon degradation, distribution shift
- Evaluate the 2026 research landscape and emerging developments in world models
- Understand the LLM to multimodal to spatial to world model to embodied AI evolution
- Assess enterprise implications and applications of world model technology
Prerequisites
- Strong understanding of AI/ML fundamentals
- Familiarity with deep learning and neural networks
- Understanding of reinforcement learning concepts
- Basic knowledge of robotics and autonomous systems
- Familiarity with digital twin and simulation concepts
World Model vs LLM vs VLM vs Digital Twin vs Simulator
Understanding the distinctions between these concepts is critical for enterprise AI strategy. Each serves a different purpose and has different capabilities. The key distinction: Digital Twin typically represents a specific physical asset, system, process, or environment and may synchronize with real-world data. World Model typically aims to learn or represent environment dynamics in a way that supports prediction, planning, and interaction. Implementations can overlap, but they serve different primary purposes.
AI Model Type Comparison
| Model Type | What It Models | Key Capability | Enterprise Maturity |
|---|---|---|---|
| LLM (Large Language Model) | Language and text | Text generation, reasoning, Q&A | Production Ready |
| VLM (Vision-Language Model) | Vision + language | Image understanding, visual Q&A | Enterprise Adoption |
| Foundation Model | General-purpose (multi-modal) | Transfer learning, general intelligence | Enterprise Adoption |
| World Model | Environment dynamics | Predict future states, plan actions | Research/Emerging |
| Digital Twin | Specific physical asset | Real-time sync, lifecycle management | Enterprise Adoption |
| Simulator | Physics and behavior | Explicit physics modeling | Production Ready |
World Model Architecture: Perception to Action
The world model architecture follows a perception-prediction-planning-action cycle. Perception inputs (sensor data, visual inputs, telemetry, text) are encoded into a state representation (latent state, environment encoding). The world model (dynamics model, transition model, reward/prediction) uses this state to predict future states through multi-step rollout, Monte Carlo, and planning. The system can simulate actions and scenarios (what-if analysis, counterfactual reasoning) and use planning (search, optimization, policy selection) to determine actions. The action/decision (physical action, recommendation, output) is executed, and new observations update the model.
World Model Components
| Component | Function | Technology | Maturity |
|---|---|---|---|
| Perception Inputs | Sensor data, visual inputs, telemetry, text | Cameras, sensors, encoders | Production Ready |
| State Representation | Latent state, environment encoding | VAE, VQ-VAE, transformers | Research |
| World Model (Dynamics) | Dynamics model, transition model, prediction | RNN, transformer, diffusion | Research |
| Predict Future States | Multi-step rollout, Monte Carlo, planning | Model-based RL, planning algorithms | Research |
| Simulate Actions | What-if analysis, counterfactual reasoning | Generative models, simulation | Research |
| Planning | Search, optimization, policy selection | MCTS, model predictive control | Research |
| Action/Decision | Physical action, recommendation, output | Robotics, control systems | Early Adoption |
| Observe & Update | New sensor data, model refinement | Online learning, adaptation | Research |
World Models vs Digital Twins: Extensive Comparison
World models and digital twins can overlap but serve different primary purposes. A digital twin typically represents a specific physical asset, system, process, or environment and may synchronize with real-world data. A world model typically aims to learn or represent environment dynamics in a way that supports prediction, planning, and interaction. Digital twins focus on representation + synchronization + lifecycle. World models focus on dynamics + prediction + planning. Digital twins are typically explicitly modeled. World models can be learned. Both can support prediction, but world models are more focused on generative prediction and planning for AI agents.
World Models vs Digital Twins
| Aspect | Digital Twin | World Model |
|---|---|---|
| Purpose | Represent specific physical asset | Learn environment dynamics for prediction/planning |
| Input Data | Real-time sensor data from specific asset | Multi-modal: text, images, video, sensor data |
| Representation | Explicit 3D + data model | Learned latent representation |
| Physics | Explicit physics models | Learned dynamics (may or may not be physics-based) |
| Real-time Sync | Yes, with specific physical asset | May or may not sync with real-world |
| Prediction | Simulation-based prediction | Generative prediction of future states |
| Planning | Limited planning support | Core capability for action planning |
| Autonomy | Limited (mostly monitoring) | Core capability for autonomous action |
| Environment Scope | Specific asset or system | General environments |
| Enterprise Applications | Manufacturing, energy, cities | Robotics, autonomous vehicles, embodied AI |
| Maturity | Enterprise Adoption | Research/Emerging |
Architecture
The world model architecture follows a perception-prediction-planning-action cycle with continuous feedback.
1
Perception Inputs
Sensor data, visual inputs, telemetry, text from the environment. Multi-modal inputs that the world model must understand.
2
State Representation
Latent state, environment encoding. The internal representation of the current state of the world.
3
World Model (Dynamics)
Dynamics model, transition model, reward/prediction. The learned model of how the environment changes over time.
4
Predict Future States
Multi-step rollout, Monte Carlo, planning. Using the world model to predict what will happen in the future.
5
Simulate Actions/Scenarios
What-if analysis, counterfactual reasoning. Simulating different actions and their consequences.
6
Planning
Search, optimization, policy selection. Using predictions to plan optimal actions.
7
Action/Decision
Physical action, recommendation, output. Executing the planned action in the environment.
8
Observe & Update
New sensor data, model refinement. Observing the result of actions and updating the world model. Closed-loop feedback.
World Model Learning and Deployment Pipeline
From data to deployed world model
1
DATA COLLECTION: Gather environment data (video, sensors, interactions)
2
STATE ENCODING: Learn latent state representation
3
DYNAMICS LEARNING: Learn transition model (how states change)
4
PREDICTION TRAINING: Train to predict future states
5
PLANNING INTEGRATION: Integrate with planning algorithms
6
VALIDATION: Test prediction accuracy and planning performance
7
DEPLOYMENT: Deploy for robotics, autonomous systems, or simulation
8
CONTINUOUS LEARNING: Update with new observations
Model-Based Reinforcement Learning Loop
The closed-loop world model RL cycle
1
OBSERVE: Agent observes current state
2
PREDICT: World model predicts future states for candidate actions
3
PLAN: Planning algorithm selects best action
4
ACT: Agent executes action in environment
5
OBSERVE RESULT: Observe new state and reward
6
UPDATE MODEL: Update world model with new experience
7
REPEAT: Continue learning and planning
World Model Maturity Model
| Level | Name | Capability | Maturity |
|---|---|---|---|
| 1 | Static Representation | Fixed environment model, no dynamics | Production Ready |
| 2 | Predictive Environment Model | Can predict future states from current state | Research |
| 3 | Interactive Simulation | Can simulate actions and their effects | Research |
| 4 | Planning Model | Can plan actions to achieve goals | Research/Emerging |
| 5 | Agentic World Model | Can autonomously plan and act using world model | Experimental |
World Model Validation Framework
| Metric | Description | Importance |
|---|---|---|
| Prediction Error | Accuracy of predicted future states vs actual | Critical |
| Long-Horizon Degradation | How prediction accuracy degrades over time | Critical |
| Distribution Shift | Performance under different conditions than training | High |
| Rare Events | Ability to predict unusual but important events | High |
| Counterfactual Validity | Accuracy of what-if scenario predictions | High |
| Physics Consistency | Whether predictions respect physical laws | Critical |
| Temporal Consistency | Whether predictions are consistent over time | High |
| Spatial Consistency | Whether predictions are spatially coherent | High |
Enterprise Use Cases
Robotics
Research
Robot Learning
Robots use world models to predict outcomes of actions, plan movements, and learn from simulation. Sim-to-real transfer for manipulation and navigation.
Automotive
Research
Autonomous Vehicles
AVs use world models to predict traffic behavior, plan trajectories, and handle edge cases. Simulation-to-real transfer for AV training.
Manufacturing
Research
Predictive Operations
World models for factory operations prediction, what-if analysis, and autonomous optimization.
Healthcare
Research
Patient Modeling
World models for patient-specific disease progression prediction and treatment planning.
Energy
Research
Grid Prediction
World models for energy grid dynamics, demand prediction, and optimization.
Smart Cities
Research
Urban Dynamics
World models for traffic, population, and infrastructure dynamics prediction.
Enterprise AI
Research
Enterprise World Models
World models that unify structured, unstructured, and conversational data for compliant, brand-aligned AI.
AI Safety
Research
Safety Testing
World models for testing AI safety in simulated environments before real-world deployment.
Case Studies
Biosphere 2 Digital Twin and World Models (University of Arizona)
Problem: Creating a scientifically rigorous virtual replica of Biosphere 2's diverse biomes for understanding complex living systems.
Architecture: NVIDIA Omniverse for industrial metaverse, NVIDIA Cosmos for world foundation models, Unreal Engine 5 for photorealistic visualization, LiDAR & photogrammetry for 3D capture, Industrial IoT sensors for real-time monitoring.
Outcome: An unprecedented digital twin of a complex living system enabling research into Earth system responses, sustainable technology, and complex systems understanding. Students work on frontiers of AI, XR, and robotics.
Lessons: World models can enable predictive capabilities for complex environments. Multiple technologies must converge for comprehensive digital twins. Real-time sensor integration is critical.
Enterprise World Models Research
Problem: Enterprises face the dual challenge of ensuring knowledge retrieval accuracy and alignment with legal compliance and brand identity when deploying LLMs.
Architecture: Hybrid fine-tuning and context-orchestration frameworks for constructing enterprise world models that unify structured, unstructured, and conversational data.
Outcome: Research finding that generic LLMs are trained on public data but do not represent enterprise-specific behavior. Context-aware retrieval methods address knowledge accuracy. Fine-tuning addresses brand alignment. Hybrid frameworks are steps toward enterprise world models.
Lessons: Enterprise world models are an emerging research direction. Generic LLMs need enhancement for enterprise use. Hybrid approaches combining fine-tuning and context orchestration are promising.
Implementation Steps
1
Research Assessment
Understand world model capabilities and limitations. Identify potential use cases where prediction and planning are critical. Assess data availability for world model training.
2
Data Collection
Gather diverse multi-modal data: video, sensor data, interaction logs, environmental data. Ensure data quality and diversity for robust world model learning.
3
Model Development
Develop or adapt world model architecture. Train state representation, dynamics model, and prediction capabilities. Integrate with planning algorithms.
4
Validation
Validate prediction accuracy, long-horizon degradation, physics consistency, and counterfactual validity. Test under distribution shift and rare events.
5
Deployment
Deploy for specific use case (robotics, AV, simulation). Implement continuous learning. Monitor performance and safety.
Enterprise SOPs
SOP: World Model Development
Purpose: Standardize world model development process
Owner: AI Research Team
Steps: Data Collection → Architecture Design → Training → Validation → Testing → Deployment → Monitoring
Evidence: Model documentation, validation reports, performance metrics
SOP: World Model Validation
Purpose: Ensure world model predictions are accurate and reliable
Owner: AI Safety Team
Steps: Prediction Error Testing → Long-Horizon Testing → Distribution Shift Testing → Physics Consistency Check → Safety Validation
Evidence: Validation reports, safety assessments, performance benchmarks
SOP: World Model Safety
Purpose: Ensure world model-based systems are safe for deployment
Owner: AI Safety Team
Steps: Risk Assessment → Simulation Testing → Human-in-the-Loop → Gradual Deployment → Continuous Monitoring → Emergency Stop
Evidence: Safety analysis, test results, monitoring data
Lab: Create a World Model Concept
Problem: A robotics company needs to develop a world model for a warehouse robot that can predict the outcomes of its actions, plan efficient paths, and handle unexpected obstacles.
Requirements:
- Design the world model architecture
- Define state representation approach
- Specify dynamics model for environment
- Design prediction and planning integration
- Define validation framework
- Plan data collection strategy
- Design safety constraints
- Create implementation roadmap
Architecture: Perception → State Representation → World Model → Predict Future → Plan → Act → Observe → Update
Outcome: World model concept design with architecture, validation framework, and implementation roadmap for warehouse robotics.
GCC Applications
- World model research and development for enterprise AI
- Robotics world model development for industrial applications
- Autonomous vehicle world model research
- Enterprise world model development for brand-aligned AI
- AI safety testing using world models
- Synthetic environment generation using world models
- Sim-to-real transfer methodology development
- Physical AI research and development
Key Metrics & KPIs
Prediction Accuracy
Accuracy of predicted future states vs actual outcomes
Long-Horizon Performance
How well the model predicts over extended time horizons
Planning Success Rate
Percentage of plans that achieve their goals
Sim-to-Real Transfer
How well models trained in simulation perform in reality
Physics Consistency
Degree to which predictions respect physical laws
Generalization
Ability to perform in new, unseen environments
Computational Efficiency
Speed and resource requirements for prediction and planning
Safety Score
Safety performance of world model-based autonomous systems
Rare Event Prediction
Ability to predict unusual but important events
Counterfactual Accuracy
Accuracy of what-if scenario predictions
Risks & Mitigation
Prediction Error
Mitigation: Rigorous validation, uncertainty quantification, human oversight for critical decisions
Long-Horizon Degradation
Mitigation: Regular model updates, shorter planning horizons, ensemble methods
Distribution Shift
Mitigation: Domain adaptation, continuous learning, robust training, monitoring
Physics Inconsistency
Mitigation: Physics-informed training, validation against physical laws, hybrid approaches
Rare Event Failure
Mitigation: Scenario testing, edge case generation, safety constraints
Autonomous Action Safety
Mitigation: Human-in-the-loop, safety constraints, gradual autonomy, emergency stops
Maturity Model
1
Static Representation
Fixed environment model with no dynamics. Simple state representation.
2
Predictive Environment Model
Can predict future states from current state and action. Basic dynamics learning.
3
Interactive Simulation
Can simulate actions and their effects on the environment. What-if analysis.
4
Planning Model
Can plan actions to achieve goals using world model predictions. Model-based RL.
5
Agentic World Model
Can autonomously plan and act using world model. Embodied AI with world model reasoning.
2026 Trends & Emerging Developments
World Foundation Models (NVIDIA Cosmos)
Research
Generative World Models
Research
Video Prediction Models
Research
3D World Models
Research
Embodied AI with World Models
Emerging
VLA (Vision-Language-Action) Models
Research
Model-Based Reinforcement Learning
Research
Sim-to-Real with World Models
Emerging
Enterprise World Models
Research
Physical AI with World Models
Experimental
Career Applications
World Model EngineerPhysical AI EngineerRobotics AI ResearcherAutonomous Systems EngineerModel-Based RL SpecialistEmbodied AI ResearcherAI Safety ResearcherSim-to-Real Transfer EngineerGenerative World Model ResearcherEnterprise World Model ArchitectPhysical AI ArchitectAutonomous Vehicle AI ResearcherSpatial AI EngineerAI Planning EngineerCognitive AI Researcher
Frequently Asked Questions
Q: What is a world model?
A: A world model is a learned representation of an environment, its dynamics, and possible future states that supports prediction, planning, or action. It is distinct from digital twins, LLMs, and simulators.
Q: How is a world model different from an LLM?
A: An LLM models language and text. A world model models environment dynamics and physical world behavior. LLMs generate text; world models predict physical world states.
Q: How is a world model different from a digital twin?
A: A digital twin represents a specific physical asset with real-time synchronization. A world model learns environment dynamics for prediction and planning. They can overlap but serve different primary purposes.
Q: How is a world model different from a simulator?
A: A simulator explicitly models physics using mathematical equations. A world model learns dynamics from data. Simulators are explicit; world models are learned.
Q: What is model-based reinforcement learning?
A: Model-based RL uses a learned world model to predict outcomes of actions, enabling more efficient planning and learning compared to model-free RL.
Q: What are latent dynamics models?
A: Latent dynamics models learn to represent environment state in a compressed latent space and model how that latent state evolves over time.
Q: What are generative world models?
A: Generative world models use generative AI techniques (like diffusion or VAEs) to generate predictions of future world states, including video prediction.
Q: What is video prediction in world models?
A: Video prediction is the task of predicting future video frames from past frames, a form of world modeling for visual environments.
Q: What are 3D world models?
A: 3D world models learn to represent and predict 3D environments, enabling spatial reasoning and planning for robotics and autonomous systems.
Q: What is embodied AI?
A: Embodied AI is AI that has a physical embodiment (robot, vehicle) and can perceive, reason, and act in physical environments, often using world models.
Q: What are VLA (Vision-Language-Action) models?
A: VLA models combine vision, language, and action capabilities, enabling AI systems that can see, understand instructions, and take physical actions.
Q: What are world foundation models?
A: World foundation models are large-scale learned representations of environment dynamics that can support prediction, planning, and reasoning across diverse physical environments. NVIDIA Cosmos is an example.
Q: How do you validate a world model?
A: Validation includes prediction error analysis, long-horizon degradation assessment, distribution shift testing, counterfactual validity, physics consistency, and temporal/spatial consistency.
Q: What is long-horizon degradation?
A: Long-horizon degradation is how prediction accuracy decreases over longer prediction timeframes, a key challenge for world models.
Q: What is counterfactual reasoning in world models?
A: Counterfactual reasoning is the ability to predict what would have happened if a different action had been taken, important for planning and learning.
Q: How do world models support robotics?
A: World models enable robots to predict outcomes of actions, plan movements, learn from simulation, and handle unexpected situations.
Q: How do world models support autonomous vehicles?
A: AVs use world models to predict traffic behavior, plan trajectories, anticipate pedestrian movements, and handle edge cases.
Q: What is the relationship between world models and sim-to-real transfer?
A: World models can serve as learned simulators for training AI systems, with sim-to-real transfer deploying the trained systems to real environments.
Q: What are enterprise world models?
A: Enterprise world models are an emerging research direction that unifies structured, unstructured, and conversational data for compliant, brand-aligned enterprise AI.
Q: What is NVIDIA Cosmos?
A: NVIDIA Cosmos is a world foundation model platform for building world models that can understand and predict physical world dynamics.
Q: How do world models relate to AI safety?
A: World models enable testing AI systems in simulated environments before real-world deployment, improving AI safety through scenario testing and validation.
Q: What is the LLM to world model evolution?
A: The evolution goes: LLM → Multimodal Model → Spatial Model → World Model → Embodied/Physical AI, with each stage adding capabilities for understanding and interacting with the physical world.
Q: Are world models production-ready?
A: Most world model applications are still in research or early experimental stages. Production deployment requires careful validation, especially for safety-critical applications.
Q: What is the reality gap for world models?
A: The reality gap is the difference between world model predictions and real-world behavior, caused by incomplete learning, distribution shift, and unmodeled dynamics.
Q: How do world models handle uncertainty?
A: World models can incorporate uncertainty through probabilistic predictions, ensemble methods, and uncertainty quantification.
Q: What is the relationship between world models and digital twins?
A: Digital twins provide real-world data that can train world models. World models can enhance digital twins with predictive and planning capabilities. They are complementary.
Q: How do world models support planning?
A: World models enable model-based planning by predicting outcomes of candidate actions, allowing the system to search for optimal action sequences.
Q: What is the role of world models in Physical AI?
A: World models are a core component of Physical AI, enabling AI systems to understand physical environments, predict consequences of actions, and plan physical actions.
Q: What are the limitations of current world models?
A: Limitations include prediction accuracy, long-horizon degradation, distribution shift, physics consistency, computational cost, and the need for large training datasets.
Q: How do you train a world model?
A: World models are trained on environment data (video, sensor data, interaction logs) using self-supervised learning, reinforcement learning, or supervised learning approaches.
Q: What is the future of world models?
A: The future includes world foundation models, better sim-to-real transfer, enterprise world models, embodied AI, and convergence with digital twins and physical AI.
Q: How do world models relate to the industrial metaverse?
A: World models can power the industrial metaverse by providing realistic environment simulation and prediction for virtual factories and digital twins.
Q: What is the Biosphere 3 project?
A: Biosphere 3 is the University of Arizona's digital twin of Biosphere 2, using NVIDIA Omniverse and Cosmos world foundation models for complex living system research.
Q: How do world models support decision-making?
A: World models support decision-making by predicting outcomes of different choices, enabling what-if analysis, and recommending optimal actions.
Q: What is the relationship between world models and foundation models?
A: World foundation models are a type of foundation model specifically trained to understand and predict physical world dynamics, enabling transfer learning for physical AI applications.
Q: How do world models handle multi-modal inputs?
A: World models can process multi-modal inputs (text, images, video, sensor data) through multi-modal encoders that create unified state representations.
Q: What is the role of world models in autonomous manufacturing?
A: World models can predict factory operations, optimize production, handle disruptions, and enable more autonomous manufacturing decisions.
Q: How do you ensure world model safety?
A: Safety is ensured through rigorous validation, human-in-the-loop for critical decisions, safety constraints, gradual deployment, and continuous monitoring.
Q: What is the relationship between world models and AGI?
A: World models are considered a potential path toward AGI, as they enable AI to understand and interact with the physical world, a key capability for general intelligence.
Q: How do world models support AI agents?
A: World models enable AI agents to predict consequences of their actions, plan effectively, and make informed decisions in complex environments.
Q: What is the enterprise potential of world models?
A: Enterprise potential includes predictive operations, autonomous decision-making, enhanced simulation, AI safety testing, and brand-aligned enterprise AI.
Research References
NVIDIA Cosmos (2026). "World Foundation Models for physical AI and robotics." [Industry]
University of Arizona (2026). "AI and World Models: Biosphere 3 Digital Twin Project." [Academic]
World Economic Forum (2024). "Technological Convergence and World Models." [Industry]
Enterprise World Models Research (2026). "Hybrid fine-tuning and context-orchestration for enterprise world models." [Research]
Physical Reasoning Benchmarks (2026). "Evaluation of world model physics adherence (Wan2.2, etc.)." [Research]
IEEE/ACM Research Papers (2026). "World model architectures, model-based RL, and embodied AI." [Research]