World Action Models Are Having a Moment, But Haven't We Been Here Before?

Meanwhile, PointAction takes a different approach entirely. Instead of working purely in RGB space, they predict future 3D point clouds, what they call "dynamic 3D pointmaps." The argument is that video predictions are fundamentally ambiguous about metric distances and contact geometry (which is true), and that explicit 3D representations make action grounding less fuzzy. They report state-of-the-art results on robot manipulation tasks, though I'd note the experiments are still relatively constrained.

WAM-Nav pushes into navigation specifically, using what they call "asymmetric joint diffusion" to generate both actions and visual foresight simultaneously. They're claiming a 15.7% improvement in success rates on image-goal navigation tasks and an 85% success rate in real-world deployment. Those are impressive numbers if they hold up!

What nobody's talking about

Here's where I get skeptical, call me old-fashioned. Every one of these papers is optimizing for benchmark performance on tasks that are, frankly, still pretty simple compared to real-world deployment scenarios. Reaching for objects on a table. Navigating cluttered rooms. Following a human through a space.

These are hard problems! I'm not dismissing them. But the implicit assumption running through all this work is that the same approaches will scale to genuinely complex, open-ended environments. And that assumption is doing a lot of heavy lifting that nobody seems interested in examining.

I've seen this movie before. In 2016, everyone was convinced that end-to-end learning would solve self-driving cars within a few years. The benchmarks looked great. The demos were impressive. And then the field spent the next eight years discovering all the edge cases that benchmarks don't capture.

The other thing that bothers me is the resource question. CLAW is particularly interesting because it learns from action-free videos, which theoretically means you could train on YouTube or whatever massive video dataset you have lying around. But training these models still requires substantial compute, and the papers are notably quiet about actual training costs. When I see phrases like "extensive experiments across diverse tasks and embodiments" without corresponding cost breakdowns, I start to wonder what's being left out.

The 3D question

One genuine point of progress, and I'll give credit where it's due, is the move toward explicit 3D reasoning. 3DThinkVLA is doing something genuinely clever here, injecting 3D geometric priors into vision-language-action models without requiring 3D sensors at deployment time. They're basically using 3D foundation models as teachers during training, then discarding them once the student model has learned the relevant spatial intuitions.

This matters because the dirty secret of a lot of robotics research is that it quietly assumes access to depth sensors or structured environments that don't exist in the real world. If you can get 3D reasoning from 2D cameras alone, that's a meaningful step toward deployable systems.

UnsOcc is tackling a related problem from a different angle, focusing on "unstructured scenes" like open-pit mines where traditional perception methods fall apart. They built their own dataset (which is admirable, more researchers should do this instead of endlessly recycling the same benchmarks) and developed a multi-modal fusion approach that handles the weird distribution of obstacles you get in mining environments.

But here's the thing, we don't know yet whether these 3D approaches will generalize across the wild variety of real-world environments. The experiments are promising, but they're also carefully scoped. It's too early to say whether we're looking at genuine breakthroughs or incremental improvements that will plateau once you throw genuinely novel scenarios at them.

So what

Look, I'm not trying to be the old guy yelling at clouds here. The technical work is solid, the researchers are clearly talented, and the underlying intuition (robots should imagine before they act) is probably correct.

What concerns me is the narrative forming around this work. The framing in these papers, and increasingly in the coverage of them, suggests we're on a clear path toward general-purpose robotic intelligence. That framing is premature. We've got a collection of methods that work well on specific benchmarks under specific conditions, and we've got a lot of assumptions about scaling that haven't been tested.

The field would benefit from more papers that focus on failure modes. More work on edge cases. More honest accounting of computational costs and deployment constraints. More skepticism about whether benchmark performance actually predicts real-world capability.

Maybe I'm wrong! These young founders and researchers might crack it faster than I expect. But I've been watching tech cycles for thirty years now, and the pattern of early optimism followed by a long slog through unexpected complications is remarkably consistent.

If you want to argue about this, my email's on the about page. I still prefer email to Slack, and I actually read the responses.

Sarah Williams · 8 hours ago · 5 min