AI xPress

AI and computer science research.

Back to Foragentis

Robotics & Embodied AI

WorldArena benchmark finds high video quality does not mean strong task skills in embodied world models

Robotics & Embodied AI November 21, 2026

A robot arm welding a car frame

In a 2026 paper in AI Open, researchers report a gap between how good an embodied world model's predicted video looks and how useful that model is for tasks. Embodied world models are AI systems that predict how an action changes an environment. The team created WorldArena, a benchmark that scores models on video perception quality and on embodied task functionality. Video perception quality uses 16 metrics across six sub-dimensions. Embodied task functionality tests the models as data engines, policy evaluators, and action planners, and includes human evaluation. Across 14 representative models, high visual quality did not reliably translate into strong task ability. The researchers call this a perception-functionality gap. They also propose EWMScore, a combined metric that brings the multiple performance measures into one interpretable index, and release a public leaderboard.

What does WorldArena measure?

WorldArena measures two things at once. First, it checks video perception quality with 16 metrics across six sub-dimensions. Second, it checks embodied task functionality by testing world models as data engines, policy evaluators, and action planners, and by adding human evaluation. The design is meant to show whether a model that looks good in video is also useful for later decisions.

How does visual quality relate to task skill?

Across 14 representative models, the two did not move together. A model could score high on video quality and still show weak embodied task ability. The researchers describe this as a significant perception-functionality gap. That finding is the reason they argue that judging world models mainly by video quality misses part of what matters.

What else should I know?

Does this benchmark connect to reinforcement learning and robotics?

WorldArena belongs to embodied AI, the area where agents act in environments. Its functional tests include policy evaluation and action planning, which are roles that often matter in reinforcement learning and robotics.

What does EWMScore add?

EWMScore is a combined metric that integrates the multiple performance measures into a single interpretable index. It gives one overall way to compare a model's standing.

Where are the results?

The abstract says the WorldArena benchmark is released with a public leaderboard, so the model results are available publicly.

Limits and what's next

The abstract describes a benchmark study of 14 representative embodied world models. The study evaluates these models with the WorldArena benchmark. The design combines metrics with human evaluation. The abstract does not say how these models would perform outside the benchmark. The researchers frame the public leaderboard as a way to track progress toward world models that are functional in embodied AI.

Source: Yu Shang et al., “WorldArena: A unified benchmark for evaluating perception and functional utility of embodied world models,” AI Open, 2026.