Static guide overview for crawlers and no-JS readers.
| Attribute | Value |
|---|---|
| Guide | World Models for Robotics |
| Summary | How to use world models for robot learning: from simulation-based training to real-world deployment and sim-to-real transfer. |
| Related Models | 5 |
| Related Research | 2 |
| References | 3 |
Editorial guide section preserved directly in static HTML.
Robots operate in the real world where every trial is expensive, slow, and potentially destructive. World models let robots learn primarily in imagination, predicting outcomes of actions before executing them physically. This dramatically reduces the number of real-world interactions needed and enables safe exploration.
Editorial guide section preserved directly in static HTML.
The classic robotics pipeline trains in a hand-crafted simulator, then transfers to reality. World models improve this: they can learn from real-world data to build more accurate simulators, fine-tune pre-trained policies with fewer real interactions, and adapt to real-world dynamics that hand-crafted simulators miss.
Editorial guide section preserved directly in static HTML.
TD-MPC2 excels at multi-task manipulation with 104 tasks from a single model. DreamerV3 provides general-purpose sample-efficient control. UniSim acts as a learned universal simulator for diverse robotic scenarios. NVIDIA Cosmos offers industrial-scale world simulation for autonomous vehicles and robots.
Editorial guide section preserved directly in static HTML.
For robotic manipulation, world models predict how objects respond to grasping, pushing, and assembly actions. TD-MPC2 and DreamerV3 have shown strong results in simulated manipulation benchmarks. The challenge remains contact-rich interactions where dynamics are highly discontinuous.
Editorial guide section preserved directly in static HTML.
World models for locomotion predict terrain dynamics, body states, and energy expenditure. They enable robots to plan efficient movement strategies by imagining different gaits and trajectories. DreamerV3 has demonstrated locomotion learning across multiple body types in DMControl.
Editorial guide section preserved directly in static HTML.
The next frontier is foundation world models for robotics, pre-trained on massive amounts of real-world interaction data, then fine-tuned for specific robots and tasks. NVIDIA Cosmos and AMI World Model represent early steps toward this vision.
| Model | Lab | Category | Year |
|---|---|---|---|
| TD-MPC2 | MIT / Meta | Model-Based RL | 2024 |
| DreamerV3 | Google DeepMind | Model-Based RL | 2023 |
| NVIDIA Cosmos | NVIDIA | Foundation World Model | 2024 |
| UniSim | Google DeepMind | Generative World Model | 2023 |
| AMI World Model | AMI Labs | Foundation World Model | 2024 |
| Topic | Summary |
|---|---|
| World Models for Robotics | How world models improve robot learning, learned simulation, safe exploration, and sim-to-real transfer across manipulation, navigation, and control. |
| Model-Based Reinforcement Learning | What model-based reinforcement learning is, how world models enable imagination-based planning, and why Dreamer, MuZero, PlaNet, and TD-MPC2 matter. |
FAQ answers rendered directly into static HTML for extractable responses.
For general manipulation: TD-MPC2. For sample-efficient learning from scratch: DreamerV3. For industrial-scale simulation: NVIDIA Cosmos. For universal simulation from diverse data: UniSim.
Combine world model training on real data with domain randomization in learned simulators. Fine-tune policies with a small amount of real-world interaction. Use uncertainty-aware world models that know when their predictions are unreliable.
Short extractable summary preserved directly in static HTML.
Editorial provenance and refresh policy preserved directly in static HTML.
Published by world-models.io editorial board.
Lead editor Bernard Grenat.
This guide page turns primary-source material into structured workflows, explanations, and linked reading paths.
Each editorial page is assembled from primary sources, normalized into extractable summaries, checked for factual drift, and reviewed before publication or major refreshes. Last reviewed: 2026-06-21.
Pages are refreshed when a new paper, benchmark, release, architecture update, or stronger primary source materially changes the answer a reader or AI system should retrieve.
Each page links back to relevant primary sources and keeps a stable canonical URL so readers can verify claims, trace context, and reference the most up-to-date version. See the editorial policy.