I don't quite understand Cangqiong CEORION's zero-shot number

Just saw Seahi Intelligence's Cangqiong CEORION, an ocean embodied large model, focused on autonomous planning under weak or even no communication. Honestly I'm more curious about how it stuffs a brain into the water.

The most striking detail in the material: in simulation tests, task success rate exceeds 90%, and the positioning success rate for fine grasping also exceeds 90%, said to reach the level of professional diver operators. There's another number: facing unseen sea areas, water quality, lighting, and different robot platforms, zero-shot adaptation capability exceeds 70%.

I stared at that 70% for quite a while.

I've grinded LeetCode for two months and done a bit of RAG work. A model holding onto 70% on unseen data is already not low, especially underwater where even light can't penetrate. But the fault tolerance for grasping is different from doing problems—getting a problem wrong once is a WA, getting it wrong once underwater might be a collision. Its promo says it reduces underwater collision accident rate by 80%; putting that together with 70% zero-shot adaptation, I'd want to follow up with one question: that 30% of non-adaptation, which working conditions does it fall on.

Not being contrarian. I've been preparing for algorithm jobs lately, and I increasingly feel that metrics like success rate depend on who sets the denominator.

Robots equipped with Cangqiong CEORION can cover 12 categories of underwater operation scenarios including inspection, detection, cleaning, and grasping, without frequently switching models.

If this generality is real, then its value isn't in how big the model itself is, but in how many specialized models it saves you from training. On this point I agree.

Recommended for people doing embodied AI, ocean engineering, or multimodal deployment to take a look at the technical disclosure; those purely here for leaderboard chasing can wait a bit longer.