HTTP/2 200 server: GitHub.com content-type: text/html; charset=utf-8 last-modified: Tue, 17 Jun 2025 07:00:54 GMT access-control-allow-origin: * strict-transport-security: max-age=31556952 etag: W/"685112a6-58a5" expires: Sun, 28 Dec 2025 08:09:01 GMT cache-control: max-age=600 content-encoding: gzip x-proxy-cache: MISS x-github-request-id: 9F88:1F53DD:762B66:846942:6950E33E accept-ranges: bytes age: 0 date: Sun, 28 Dec 2025 10:01:40 GMT via: 1.1 varnish x-served-by: cache-bom-vanm7210078-BOM x-cache: HIT x-cache-hits: 0 x-timer: S1766916101.596087,VS0,VE201 vary: Accept-Encoding x-fastly-request-id: 2f7e861c36e05851d9561319f87cf5b77ce050fd content-length: 3296

AdaWorld

Learning Adaptable World Models with Latent Actions

Shenyuan Gao¹, Siyuan Zhou¹, Yilun Du², Jun Zhang¹, Chuang Gan^3,4

HKUST¹, Harvard², UMass Amherst³, MIT-IBM Watson AI Lab⁴

ICML 2025

Paper Code Model

TL;DR: AdaWorld is a highly adaptable world model pretrained with continuous latent actions from thousands of environments, enabling zero-shot action transfer, fast adaptation, and new skill acquisition with minimal finetuning.

Select one set below to view different action transfer results.

Robot #1 Robot #2 Robot #3

source video → target scene

source video → target scene

source video → target scene

source video → target scene

source video → target scene

source video → target scene

Human #1 Human #2 Human #3

source video → target scene

source video → target scene

source video → target scene

source video → target scene

source video → target scene

source video → target scene

Game #1 Game #2 Game #3

source video → target scene

source video → target scene

source video → target scene

source video → target scene

source video → target scene

source video → target scene

AdaWorld enables more effective planning.

On Procgen: [Left] action-agnostic baseline; [Right] AdaWorld (ours).

On Robosuite: [Left] goal image; [Mid] action-agnostic baseline; [Right] AdaWorld (ours).

---

Latent Action Autoencoder

With an information bottleneck design, our latent action autoencoder is able to extract the most critical action information from videos and compresses it into a continuous latent action.

Autoregressive World Model

We extract latent actions from unlabeled videos using the latent action encoder. By leveraging the extracted actions as a unified condition, we pretrain a world model that can perform autoregressive rollouts at inference.

---

Experiment Results

Efficient World Model Adapation to Unseen Environments

Visual Planning

Action Composition

Customizable Actions with Strong Controllability

Qualitative Comparison with Other Variants

Latent Action Visualization

---

BibTeX

@inproceedings{gao2025adaworld,
 title={AdaWorld: Learning Adaptable World Models with Latent Actions},
 author={Gao, Shenyuan and Zhou, Siyuan and Du, Yilun and Zhang, Jun and Gan, Chuang},
 booktitle={International Conference on Machine Learning (ICML)},
 year={2025}
}