OpenViSTAI · Agents · Blockchain · Crypto
Submit a project
Menu
AI Evaluation & Observability

World In World

World In World is a Python benchmark and toolkit that evaluates visual world models through closed-loop embodied tasks, using shared online planning and an action API that adapts to text, viewpoint and low-level controls.

Explore the practical guide ↓

Independent OpenViST project profile, based on linked sources.

World In World — OpenViST project profile in AI Evaluation & Observability.
OpenViST project illustration. See the linked sources for product details.
Understand the project

Practical overview

When it fits

Use it when an evaluation needs to test how a visual world model contributes to an agent’s behavior through repeated observations and actions, across one of the included task types.

Who it is for

  • Researchers evaluating visual world models for embodied AI.
  • Teams comparing models on task performance involving perception, planning and action.
  • Researchers studying active recognition, embodied question-answering, image-goal navigation or robotic manipulation.

Documented uses

  • Compare visual world models by their usefulness to an embodied agent, not only by generated image or video appearance.
  • Evaluate active recognition and embodied question-answering behavior.
  • Test image-goal navigation with a shared online-planning strategy.
  • Assess model utility in robotic manipulation tasks.
Before you choose

Requirements & limits

The public repository lists an MIT license.

  • Select the task before setup: AR, IGNav and AEQA use Habitat-sim, while manipulation uses wow-manip.
  • Dataset requirements are task-specific: AR uses MP3D; IGNav and AEQA use HM3D.
  • The run checklist calls for different policies and extra models by task; for example, AR may use a heuristic policy, while AEQA requires VLM deployment and Grounding SAM2.
  • The repository describes its full documentation and tutorials as under construction.

Check current requirements in the official documentation. Source-based guidance does not mean OpenViST has run or independently tested this project.

Source-based practical guide

Prepare a task-specific evaluation

Identify and gather the documented environment, datasets and policy components needed before running an evaluation.

Before you start

  • A task choice: Active Recognition (AR), Image-Goal Navigation (IGNav), Active Embodied QA (AEQA) or manipulation.
  • For AR, IGNav and AEQA, the Habitat-sim environment; for manipulation, wow-manip.
  • The scene datasets and evaluation episodes specified for the chosen task.

Steps

  1. Set up Habitat-sim for AR, IGNav or AEQA, or set up wow-manip for manipulation.

  2. Obtain the task’s scene dataset: MP3D for AR and HM3D for IGNav or AEQA; the manipulation environment data is identified in the repository.

  3. Download the corresponding AR, IGNav or AEQA evaluation episodes; manipulation episodes are identified in the repository.

  4. Consult the run-command instructions to determine which policy and any extra task model the chosen evaluation requires.

Expected outcome

The task-specific environment, data and required evaluation components are identified and prepared for the documented run stage.

Guide limitations

  • This preparation guide does not cover deploying policies or executing evaluation scripts.
  • Policy and additional model requirements differ by task; AR and IGNav can use a heuristic policy instead of deploying the VLM policy.
  • The repository says full documentation and tutorials are under construction.
Documentation 1 ↗Documentation 2 ↗
Evidence and transparency

Sources & review

Official repository ↗Guide documentation ↗
Source record details
Source availability
Active
Sources checked
16 October 2026
Original website title
World-In-World/world-in-world · Official homepage title · retained in its source language
How we review sourcesSuggest a correction