"World model" is getting overused these days

A few days ago I came across a post from the Beijing Humanoid Robot Innovation Center — under the TianGong open-source plan there's a WoW, called the embodied world model. I stared at that description twice: unified perception, decision-making, and prediction, empowering robots in the real world. The words are all correct, but after reading it I couldn't say what it actually delivers.

I've recently been working on a visual spec for a robot status panel, and started touching deep learning-related stuff two months ago, so I'm already a bit immune to the phrase "model capability." Whenever "unified" appears in marketing copy, I first go looking for where its boundaries are written.

I didn't find them.

I went over that intro on their official site again and again. What I could see was the embodied world model driving data augmentation generation, and a contact/cooperation entry point. What I couldn't see was what this thing takes as input, what it outputs, and under what conditions it fails. I searched around for evaluations — there is a benchmark called WoW-World-Eval, specifically testing video foundation models' performance in embodied scenarios, and the paper mentions that current video generation models produce footage that looks pretty good visually. "Pretty good" is a subtle word. People who build design systems are most wary of "pretty good," because between pretty good and usable lie empty states, error states, and extreme inputs.

Last week I wrote a post about demo state versus delivery state, and my view hasn't changed. Video demos in the robotics field and high-fidelity prototypes in design drafts are the same thing — both show you the smoothest possible path. The real value of world models like PelicanVLM should show in how it falls back when planning fails, not in generating a smooth prediction video.

My judgment is one line: it's still too early to talk about the capabilities of embodied world models. Writing the failure boundaries into the docs is more useful than putting out a few more pretty clips.

By the way, has anyone actually run it on tasks with interference, like sudden lighting changes or an occluded target? I'd really like to know how it performs then.