The VLA Gold Rush Is Here, and I've Seen This Movie Before

Another paper, GSAM, focuses on articulated objects, things with hinges and joints like doors and cabinets. They claim a 36% improvement in manipulation success rate over their best baseline, which sounds impressive until you realize we don't know what that baseline actually achieved. Still, the approach is clever: they use a large language model to generate constraint functions that prevent the robot from, you know, smashing into things. Experiments covered 50 hinge tasks across 5 object categories, which isn't huge but it's more than a single demo setup.

The data problem nobody wants to talk about

Here's where I get grumpy. Every single one of these papers depends on massive amounts of training data, and getting that data is brutally hard. The RDGen paper actually tries to address this head-on by using reinforcement learning in simulation to generate demonstrations automatically instead of relying on human teleoperation. The authors are pretty direct about the problem: human-collected data is "labor-intensive, costly, and difficult to scale." No kidding!

But what do I know. Maybe the kids building these systems have figured out something the last generation didn't. The HARP-VLA paper proposes using human videos (which are abundant on YouTube and elsewhere) as training data for robot policies. The trick is bridging what they call "cross-embodiment discrepancies," basically the fact that human hands don't look or move like robot grippers. They report a 7.1% real-world success rate improvement over their strongest baseline, which, again, remains unclear in absolute terms but suggests the approach has legs.

What's actually new versus what's hype

Let me try to separate signal from noise here, because I think there's genuine progress mixed in with the usual academic paper inflation.

The genuinely interesting development is this: researchers are finally taking the physics of manipulation seriously. The AttenA+ framework makes what seems like an obvious point, that not all moments in a robot trajectory are equally important. When you're moving fast across open space, errors are forgivable. When you're doing the delicate bit where you actually grasp something, precision matters enormously. Current training treats all these moments the same, which is, well, dumb. AttenA+ reweights the training objective based on velocity, and they claim improvements to 98.6% on something called the Libero benchmark. That's a 1.5% gain, which sounds tiny but at the high end of benchmarks, small gains are hard.

Similarly, ELAN4D adds what they call "4D supervision," basically making the model predict where the robot's joints will be in the future, not just what action to take right now. This is the sort of physics-aware reasoning that seems obviously necessary but apparently wasn't being done. They show improvements under "out-of-distribution perturbations," which is academic speak for "the robot still works when you move the camera or change the background."

The less interesting development, in my view, is the continued reliance on ever-larger language and vision models as the backbone for these systems. It works, sort of, but it feels like we're throwing compute at problems instead of actually understanding them. I've seen this pattern before too.

So what does this mean for actual robots

It's too early to say whether this wave of research will translate into robots that work in real homes and factories. The gap between benchmark performance and real-world deployment is vast, and nobody's really solved it yet. These papers are all tested in controlled settings, even when they claim "real-world" experiments.

But I'll say this: the field is asking better questions than it was five years ago. Instead of "how do we make the robot do one specific thing really well," the question is now "how do we make the robot generalize to things it hasn't seen before." That's the right question. Whether these specific approaches are the right answers, well, we won't know for another few years.

If you want to argue with me about this, my email's on the about page. I still check it.

James Chen · 6 hours ago · 4 min