1. Problem Statement

Robot learning is currently transitioning from task-specific control pipelines to foundation-model-driven embodied intelligence, primarily through Vision-Language-Action (VLA) policies. While these policies show promise in generalization, they face critical limitations in complex physical environments:

• Long-horizon Reasoning: Purely reactive VLA policies struggle with tasks requiring planning over extended timeframes.
• Temporal Credit Assignment: They often fail to attribute success or failure to specific actions in a sequence.
• Robustness: They are susceptible to compounding errors when deployed in the real world.
• Lack of Predictive Structure: Current methods often lack an explicit mechanism to anticipate how the environment will evolve under specific actions, leading to a disconnect between semantic intent and physical realizability.

The literature on World Models (predictive representations of environment dynamics) is currently fragmented across architectures (video generation, latent dynamics, symbolic models), functional roles (planning, simulation, evaluation), and application domains. There is a need for a unified, robot-centric perspective to clarify how predictive models can be effectively coupled with robot policies to overcome these limitations.

2. Methodology and Taxonomy

The authors propose a comprehensive, robot-learning-centered taxonomy that categorizes world models based on their architectural coupling with policies and their functional roles. The survey organizes the literature into three primary pillars:

A. World Models for Policy (Architectural Coupling)

This section examines how predictive models are integrated into the policy loop. The authors identify five distinct architectural paradigms:

1. Inverse Dynamics (IDM) Style (Decoupled): A two-stage pipeline where a world model (often a video generator) predicts future observations, and a separate policy module (inverse dynamics model) infers actions from the current state and the predicted future. (e.g., UniPi, VidMan).
2. Single-Backbone Style (Unified): A single generative backbone jointly models future visual evolution and action generation. Observation and action tokens are processed in a shared latent space, allowing the policy to inherit spatiotemporal priors from video pretraining. (e.g., UVA, Cosmos Policy).
3. MoE/MoT Style (Expert Fusion): Maintains separate expert streams for video prediction and action generation but couples them through shared attention mechanisms (Mixture-of-Transformers). This balances specialization with deep interaction. (e.g., Motus, LingBot-VA).
4. Unified VLA Models: Future-oriented predictive structures (e.g., future image prediction, visual foresight) are internalized within a unified VLA backbone, often using joint training objectives rather than separate modules. (e.g., GR-1, DreamVLA).
5. Latent-Space World Modeling: Avoids explicit pixel/video generation. Instead, the model learns compact, future-aware latent representations or predictive embeddings (often inspired by JEPA) to guide action generation, reducing computational overhead. (e.g., VLA-JEPA, FLARE).

B. World Models as Simulators (Functional Roles)

Beyond conditioning, world models serve as interactive environments:

1. For Reinforcement Learning (RL): The world model acts as a learned simulator (a "digital twin") where policies can be trained via imagined rollouts, replacing costly real-world interaction. This enables offline adaptation and policy improvement (e.g., World-Env, VLA-RFT).
2. For Evaluation: The world model acts as a validator, scoring candidate action sequences by simulating their outcomes. This supports decision-time selection, safety filtering, and policy ranking without physical execution (e.g., WorldEval, GPC).

C. Robotic Video World Models (Generation Capabilities)

The survey traces the evolution of video generation specifically for robotics:

• Imagination Engine: Using video generation to synthesize synthetic data or visual plans for supervision.
• Action-Controllable: Ensuring generated futures strictly follow commanded action sequences (causal alignment).
• Structure-Aware: Incorporating geometric priors (masks, depth, 3D flow) to improve physical consistency.
• Foundation-Scale: Adapting large-scale video foundation models (e.g., Cosmos, Genie) into reusable, interactive world interfaces for robotics.

3. Key Contributions

• Unified Robot-Centric Framework: The paper moves beyond generic definitions to define world models specifically in the context of embodied AI, emphasizing action-conditioned consistency and physical plausibility over mere visual realism.
• Fine-Grained Taxonomy: It provides a structured classification of world models based on architectural paradigms (IDM, Single-Backbone, MoE, etc.) and functional roles (Policy, Simulator, Evaluator), revealing nuances often overlooked in broader surveys.
• Comprehensive Resource Curation: The authors curate and analyze representative datasets (e.g., Open X-Embodiment, DROID), benchmarks (e.g., LIBERO, WorldArena), and evaluation protocols, highlighting the gap between visual realism and functional utility.
• Identification of Core Challenges: The survey pinpoints critical bottlenecks, including causal conditioning gaps (actions not influencing predictions), efficiency bottlenecks (inference latency), and the lack of multi-modal (haptic/force) integration.

4. Results and Findings

The survey synthesizes results from recent works (2023–2026) across standard benchmarks like LIBERO, RoboTwin, and CALVIN:

• Performance: Strong performance is achievable across multiple architectural paradigms. Unified VLA and MoE/MoT approaches often show superior long-horizon performance compared to purely decoupled IDM approaches.
• Long-Horizon Robustness: While many methods perform well on short-horizon tasks (Spatial/Object suites), significant performance drops occur in long-horizon suites (Goal/Long), indicating that action-grounded consistency over time remains a major challenge.
• Simulator Utility: World models used as simulators for RL (e.g., World-Env) have demonstrated the ability to improve real-world policy performance through offline training, reducing the need for physical trial-and-error.
• Evaluation Reality: Visual realism is a poor proxy for control utility. Benchmarks like WorldArena and WorldSimBench show that models must preserve action faithfulness and physical dynamics to be useful, even if the visual output is not photorealistic.

5. Significance and Future Directions

This survey is significant because it bridges the gap between generative AI (video foundation models) and robotics control, clarifying when and why world models translate into measurable gains in real robotic behavior.

Key Future Directions Identified:

• Causal Conditioning: Developing models where predictions are strictly conditioned on specific pending actions to ensure causal alignment.
• Efficiency: Moving towards latent-space models and parameter-efficient training to reduce the computational cost of video-based world models during inference.
• Multi-Modal Integration: Incorporating haptic, force, and tactile feedback to ground predictions in physical reality, moving beyond vision-only limitations.
• Hybrid Architectures: Combining learned perceptual representations with symbolic/structured world models to enable stable, long-horizon reasoning and planning.
• Standardized Evaluation: Establishing function-aware metrics that prioritize task success, policy ranking fidelity, and executability over visual fidelity scores.

The authors maintain an accompanying GitHub repository (Awesome-World-Model-for-Robotics-Policy) to continuously track the rapidly evolving field.