The VLA arms race is heating up, and I'm trying to keep track

I initially thought this was just incremental optimization work, but after reading the paper more carefully, it seems like a genuine architectural insight. If you tell a robot to "pick up the mug," it doesn't need to process every pixel of the background wall. CogVLA makes that selective attention explicit.

The code is open-sourced, which is increasingly rare as this field gets more competitive.

When language isn't enough

Here's something I hadn't considered until I read Gaze2Act: language is actually a pretty terrible way to tell a robot what to do.

Think about it. You're in your kitchen, there are three similar-looking containers on the counter, and you want the robot to hand you one specific one. "The one on the left" works until you move. "The blue one" fails if two are blue. "The one with the scratched lid" requires the robot to notice scratches.

Gaze2Act's solution is elegant. Use human eye tracking. Where you look becomes a continuous signal of intent, mapped from your first-person view to the robot's perspective through something they call "cross-view semantic matching."

They tested this on a Unitree G1 humanoid across 16 real tasks, and the results suggest gaze provides something language can't: dynamic, moment-to-moment intent that updates as the situation changes.

I should note that this requires the human to wear eye-tracking hardware, which limits practical deployment. But as AR glasses become more common, this constraint might matter less.

The small model that could

ProgVLA is doing something different. Instead of building bigger, it's asking: how small can we go while still being useful?

The answer appears to be 0.1 billion parameters. That's tiny by current standards (GPT-4 is rumored to be over a trillion). Yet ProgVLA claims competitive or better performance than much larger models, especially on long-horizon tasks.

The trick is teaching the model to understand task progress, essentially giving it an internal sense of "how far along am I?" This comes from reinforcement learning techniques, which is interesting because most VLA work has focused on imitation learning (copying human demonstrations).

I'm honestly not sure this holds up outside their benchmark conditions. Long-horizon tasks are notoriously hard to evaluate because success depends heavily on how you define the task boundaries. But if the efficiency claims are real, this matters for anyone trying to run these models on actual robot hardware rather than cloud servers.

The gap between simulation and reality

BORA addresses what might be the field's most persistent problem: robots trained in simulation fail in the real world.

Their approach is a two-phase system. First, train a critic network offline that can evaluate whether a proposed action is good or bad. Then, in the real world, use that critic to guide small corrections to the base policy, with a human in the loop to intervene when things go wrong.

The results are substantial. A 33% absolute improvement in success rate under standard conditions, and up to 43% improvement when generalizing to unseen objects. That's on dexterous manipulation tasks, which involve complex hand movements that are particularly prone to compounding errors.

What I find most interesting is the human-in-the-loop component. Rather than trying to make the robot fully autonomous, BORA assumes a human will sometimes need to step in, and builds that into the learning process. It's a more realistic deployment model than "perfect autonomy or nothing."

The representation question

RS-CL takes a different angle. The paper argues that current VLA models are using representations (the internal features they extract from images) that aren't actually well-suited for robot control.

Their solution is a contrastive loss function that aligns visual representations with proprioceptive states, basically, what the robot feels about its own body position. They claim this pushes the state-of-the-art on RoboCasa-Kitchen to 69.7% and improves real-robot success rates from 45.0% to 58.3%.

I find this paper harder to evaluate because it's more of a training technique than a new architecture. But if it's as drop-in compatible as they claim, we might see it adopted widely.

What does this mean for humanoids?

So where does this leave us? I think a few things are becoming clear.

First, the VLA paradigm is working. These models can genuinely bridge perception and action in ways that weren't possible two years ago. The benchmark numbers are getting impressive enough that the conversation is shifting from "can we do this at all" to "can we do this efficiently, reliably, and in diverse conditions."

Second, efficiency matters more than I initially appreciated. A humanoid robot that needs a data center connection to function isn't useful. The push toward smaller, faster models (CogVLA, ProgVLA) suggests the field recognizes this.

Third, we're still far from solved. The out-of-distribution numbers, even the good ones, suggest these models remain brittle when facing novel situations. The need for human-in-the-loop corrections (BORA) and alternative input modalities (Gaze2Act) points to fundamental limitations in pure language-conditioned control.

I don't know which of these approaches will win out. Tbh, it's probably some combination of all of them. But the pace of progress suggests that 2025 might be the year VLA models become actually deployable rather than just publishable.

That's a meaningful distinction. And it's one I'll be watching closely.

Sarah Williams · 2 days ago · 4 min