The VLA Arms Race Is Here, and I've Seen This Movie Before

But here's where my skepticism kicks in. The paper admits that "since VLA behavior shifts during RL, a fixed simulator can misalign with the updated policy." In other words, the robot's dream world starts diverging from reality as it learns. Their solution is to keep updating the world model alongside the policy, which makes sense in theory but adds a lot of complexity that could break in ways we don't understand yet.

The reward hacking problem nobody wants to talk about

Another paper, SOLE-R1, takes a different approach. Instead of building world models, they use a vision-language model to watch videos of the robot and judge whether it's making progress. The model does "spatiotemporal chain-of-thought reasoning," which is a fancy way of saying it thinks through what it sees frame by frame.

What caught my attention here is the admission that existing vision-language models, including GPT-5 and Gemini-3-Pro, "often fail under partial observability and distribution shift, enabling policies to exploit perceptual errors rather than solve the task." This is reward hacking, and it's a big deal. The robot figures out how to make the evaluator think it succeeded without actually doing the task. SOLE-R1 claims to be more robust to this, but the fact that the best models from OpenAI and Google can be fooled so easily should give everyone pause.

The paper says SOLE-R1 "succeeds on 24 unseen tasks" in simulation and real-robot settings. That's genuinely impressive if it holds up. But I've seen a lot of papers with impressive benchmark numbers that fell apart when someone else tried to replicate them, call me old-fashioned.

The efficiency question

Maybe the most practically interesting paper is EXPO-FT, which focuses on sample efficiency. Their headline claim: perfect task performance (30 out of 30 successes) on challenging manipulation tasks within an average of 19.1 minutes of online robot data.

Nineteen minutes! That's remarkable if true. The tasks they tested include routing string lights and plugging them in, striking a pool ball into a pocket, and inserting a flower into a wine bottle. These aren't trivial pick-and-place operations, they require precision and dynamic adjustment.

The team released an open-source codebase, which is how you know they're at least confident in their results. Nothing kills a paper faster than code that doesn't reproduce.

What the young founders are missing

Agentic-VLA introduces what they call "Adaptive Reward Synthesis," which dynamically generates reward functions based on the robot's current capabilities. It also uses a "critic model" to guide exploration rather than having the robot try random things. The benchmark improvements are substantial: +12.3% on long-horizon tasks, +28.5% in 1-shot learning.

Then there's Afford-VLA, which argues that the real problem is spatial reasoning, specifically figuring out where to interact with objects in complex scenes. Their solution involves "affordance masks" that highlight interaction regions directly aligned with action prediction. It's a more grounded approach than some of the others, literally showing the robot where to grab things.

And LACY takes yet another angle: bidirectional language-action mapping. The robot doesn't just follow instructions, it can explain what it's doing and why. The idea is that a robot that can articulate its reasoning will have better internal representations. They claim 56.46% improvement in task success rates, though the baseline matters a lot for interpreting that number.

So what

Here's the thing that bothers me about all this. Every one of these papers is solving a real problem with genuinely clever techniques. The researchers clearly know their stuff. But we've been here before, and I mean that literally.

In 2016, I wrote about deep learning for autonomous vehicles with the same breathless tone the field is using for VLAs now. The techniques were real. The improvements were real. And eight years later, we still don't have Level 5 self-driving cars. The gap between benchmark performance and real-world deployment turned out to be much wider than anyone expected.

The VLA papers all acknowledge limitations, to their credit. World-VLA-Loop admits the world model can drift. SOLE-R1 acknowledges that reward hacking remains a threat. EXPO-FT's impressive results come from specific tasks that may not generalize. But the overall narrative being constructed, that we're on the verge of robots that can learn any task quickly and reliably, feels premature.

What I'd want to see before getting excited: long-term deployment data. Not 30 trials in a lab, but thousands of hours in actual warehouses or homes, with real failure modes documented. Generalization tests across genuinely novel environments, not variations on training distributions. And honest assessments of what happens when these systems encounter situations their training never anticipated.

Until then, I'll keep reading the papers, because the technical work is legitimately interesting, while maintaining the skepticism that three decades of tech coverage has beaten into me. The kids building these systems are smart. Whether they're building something that'll change robotics or something that'll join the long list of technologies that were always "five years away," we probably won't know for another five years.

If you want to argue about any of this, my email's on the about page.

Sarah Williams · 1 hour ago · 4 min