The VLA Arms Race: Six Papers That Show Where Robot Learning Is Actually Headed

arXiv introduces DeMaVLA, focused specifically on deformable object manipulation (think: folding clothes). They argue that training separate policies for different object categories is wasteful and propose a unified foundation model pretrained on approximately 5,000 hours of real-world dual-arm demonstrations.

arXiv offers HARP-VLA, which tackles the human-robot domain gap. The insight here is that learning from human videos is promising but the visual and action representations don't transfer cleanly to robot embodiments.

arXiv presents AttenA+, arguing that the real issue is how we weight different parts of the action trajectory during training. Not all timesteps are equally important; precision-demanding moments matter more than error-tolerant transitions.

arXiv introduces GaussianDream, which adds explicit 3D spatial modeling and future prediction through a Gaussian world model plugin.

And finally, arXiv presents Wall-OSS-0.5, which asks the uncomfortable question: does VLA pretraining actually produce useful robot behavior, or is it just a better initialization for fine-tuning?

Which contributions are genuinely new?

To be precise, I'd categorize these into three tiers of novelty.

Tier 1: Asking new questions. Wall-OSS-0.5 stands out here. The field has been oddly coy about directly testing whether pretrained VLAs can do anything useful without task-specific fine-tuning. Most papers report results after fine-tuning, which makes it impossible to disentangle what the pretraining actually learned. Wall-OSS explicitly measures zero-shot real-robot behavior before any task adaptation. The answer turns out to be "yes, sort of" (they complete several tasks at high progress on a 17-task suite), which is actually a meaningful finding. After fine-tuning, they report 60.5% average task progress on 15 real-robot tasks, outperforming π₀.5 by 17.5%.

I know I'm being picky here, but this is the kind of experiment the field should have been running all along.

Tier 2: Novel technical mechanisms. ELAN4D's approach of using forward kinematics to derive 4D supervision is clever. The key insight is that you can get spatio-temporal supervision "for free" from proprioceptive states, without needing external trackers or 3D reconstruction. The auxiliary branch is discarded at inference, so there's no runtime cost. This is genuinely new in that it exploits an underutilized signal source.

AttenA+'s velocity-driven action attention is also a real contribution. The observation that low-velocity segments (precision moments) matter more than high-velocity transitions (error-tolerant motions) is obvious in retrospect, but nobody had formalized it as a training objective before. The fact that it's architecture-agnostic and adds no parameters makes it practically useful.

Tier 3: Incremental but solid. DeMaVLA, HARP-VLA, and GaussianDream all represent competent engineering rather than conceptual breakthroughs. This isn't a criticism, exactly. DeMaVLA's 5,000 hours of real-world data and human-in-the-loop correction pipeline is impressive infrastructure work. HARP's cross-embodiment alignment loss is a reasonable approach to a known problem. GaussianDream's 3D Gaussian world model is well-executed but builds directly on prior work in neural radiance fields and Gaussian splatting.

These papers will likely be useful to practitioners. They're not reshaping how we think about the problem.

What do the benchmarks actually tell us?

Here's where I get concerned. Almost all of these papers report results on LIBERO, and the numbers are getting suspiciously high:

• GaussianDream: 98.4%
• AttenA+ (improving OpenVLA-OFT): 98.6%
• ELAN4D: claims "best overall performance" (specific numbers not provided in abstract)
• DeMaVLA: "competitive performance"

When multiple methods are all claiming 98%+ on the same benchmark, we're either solving the benchmark or overfitting to it. LIBERO is useful, but it's a simulation environment with known dynamics. The gap between simulation performance and real-world deployment remains unclear.

To be fair, several papers do include real-world experiments. GaussianDream reports 50.0% on real-robot tasks. Wall-OSS reports 60.5% after fine-tuning. DeMaVLA shows results on a "household folding benchmark." But the real-world evaluations are typically on a much smaller task set, with less standardization across papers.

It's worth noting that we don't have a good shared real-world benchmark yet. Each lab tests on their own setup, making direct comparisons difficult. This isn't anyone's fault specifically, but it limits what we can conclude from the aggregate literature.

What are the methodology concerns?

A few things I'd want to see addressed:

Data scale varies wildly. DeMaVLA uses 5,000 hours of real demonstrations. Wall-OSS processes over one million trajectories per epoch. Other papers don't clearly specify their data requirements. This makes it hard to disentangle whether improvements come from the method or the data.

Compute costs are often omitted. HARP mentions "limited paired human-robot demonstrations," which is good for accessibility. But several papers don't discuss training costs at all. If a method requires 1000 GPU-hours to train, that's relevant information.

Ablation depth varies. AttenA+ provides clean ablations showing the effect of velocity weighting in isolation. Others bundle multiple contributions together, making it unclear which components drive the gains.

Real-world sample sizes are small. This is a practical constraint (real robots are expensive to run), but when you're reporting 50% success on "real-robot tasks," I'd want to know: how many trials? How much variance? What failure modes?

None of these are fatal flaws. They're the normal limitations of conference-deadline research. But they do mean we should hold the headline numbers loosely.

What does this tell us about the field's direction?

Reading these papers together, a few trends emerge.

First, the field is converging on VLA as the dominant paradigm. All six papers take for granted that vision-language-action models are the right architecture family. Nobody is questioning whether we should be doing something fundamentally different. This could be correct (VLAs might actually be the answer) or it could be a case of collective fixation on a particular approach because that's where the benchmarks and pretrained models are.

Second, there's increasing attention to what happens during training, not just architecture. AttenA+'s action attention, ELAN4D's 4D supervision, and Wall-OSS's gradient-bridged co-training all focus on the training objective rather than the model structure. This feels like a sign of maturation. The obvious architectural improvements have been picked over, so researchers are looking at subtler aspects of the learning process.

Third, the gap between simulation and reality remains the elephant in the room. Papers that report real-world results show substantially lower performance than simulation results. GaussianDream goes from 98.4% (LIBERO) to 50.0% (real robot). That's a 48 percentage point drop. We're getting better at simulation, but it's too early to say whether that translates to real-world capability.

Fourth, data is increasingly the differentiator. DeMaVLA's 5,000 hours of demonstrations, Wall-OSS's million-trajectory pretraining corpus. The labs that can collect or aggregate large-scale robot data have a structural advantage that clever algorithms alone may not overcome.

What would I want to see next?

If I were reviewing proposals in this space, I'd push for:

Standardized real-world evaluation. Something like the YCB object set but for manipulation tasks. A shared protocol that multiple labs can run independently.

Failure mode analysis. When these models fail, how do they fail? Is it perception errors, action precision, language misunderstanding, or something else? Aggregate success rates tell us very little about what's actually going wrong.

Honest compute and data accounting. Every paper should report training cost (GPU-hours or equivalent) and data requirements. This lets practitioners assess whether a method is actually accessible to them.

Longer-horizon evaluation. Most benchmarks test relatively short tasks. What happens when you chain together multiple manipulation steps? Does performance degrade gracefully or catastrophically?

Cross-paper replication. It would be valuable to see independent teams reproduce each other's results. The current situation (everyone reports their own numbers on their own setup) makes it hard to trust any individual claim.

The bottom line

Actually, the research shows that VLA models are improving, but we're still far from general-purpose manipulation. The best real-world results hover around 50-60% success, which is impressive progress but not deployment-ready. The simulation results are higher but may not transfer.

The most important contribution this week might be Wall-OSS's willingness to ask whether pretraining actually does anything useful before fine-tuning. The answer ("yes, somewhat") is encouraging, but the fact that the question hadn't been rigorously tested before suggests we've been building on assumptions rather than evidence.

For practitioners: AttenA+ seems like the easiest win, since it's architecture-agnostic and adds no parameters. ELAN4D's 4D supervision approach is also worth trying if you have access to proprioceptive data. The larger-scale efforts (DeMaVLA, Wall-OSS) require infrastructure most labs don't have.

For researchers: the low-hanging fruit is gone. The next advances will likely come from better understanding of what these models actually learn, why they fail, and how to close the sim-to-real gap. That's harder work than architecture search, but it's probably where the field needs to go.

I remain skeptical that any single paper here represents a paradigm shift. But taken together, they suggest steady progress on a genuinely hard problem. That's worth something, even if it's not as exciting as the abstracts would have you believe.

James Chen · 6 hours ago · 4 min