The Data Problem That Won't Die: Why Robot Learning Still Can't Scale

The Single-Arm Workaround

Another team took a different tack with MonoDuo. Their insight: bimanual robots are rare and expensive, but single-arm robots are everywhere. So they collect data by having a human and a robot each perform half of a two-handed task, then swap roles. Stitch the demos together synthetically, and you've got bimanual training data from monoarm hardware.

This is the kind of practical engineering hack I actually like. It's not trying to solve everything, just one specific bottleneck. The success rates aren't earth-shattering (up to 70% zero-shot, which improves to 65-70% better than training from scratch with just 25 target robot demos), but the approach is honest about its limitations.

Call me old-fashioned, but I trust incremental improvements that acknowledge their constraints more than revolutionary claims that don't.

The Vision-Free Pivot

Now here's where things get interesting from a policy perspective. A paper on VE2VF (vision-enabled to vision-free) argues that vision-based policies are fundamentally fragile. They overfit to lighting, backgrounds, camera angles, all the stuff that varies between a lab and the real world.

Their solution: train a vision-enabled teacher, then distill its knowledge into a vision-free student that relies only on pose, twist, and wrench sensing. The student achieves 95% success after roughly 50 minutes of training on the NIST assembly benchmark, with robust generalization to 8 unseen task variants.

Fifty minutes! That's fast enough to actually be practical for production deployment. And the human-in-the-loop aspect means you're not just hoping the robot figures it out.

This reminds me of what happened in computer vision around 2015, when everyone realized that the fanciest deep learning models weren't always the right choice for deployment. Sometimes simpler is better if it's more robust. The pendulum might be swinging back toward sensor-based approaches that don't require perfect visual conditions.

The VLA Arms Race

Three of the papers focus on vision-language-action models, which are basically the robot equivalent of GPT, large pretrained models that can theoretically handle any task if you describe it in natural language.

BORA tackles dexterous manipulation, which is notoriously hard because hands have so many degrees of freedom. They report a 33% absolute increase in success rate over imitation learning baselines, and up to 43% improvement on unseen object generalization. The key innovation is keeping a human in the loop during online training to catch execution errors before they break the hardware.

ProgVLA takes a different approach, building a compact 0.1B-parameter model that explicitly tracks task progress. The results are competitive with much larger models, which matters a lot for edge deployment where you can't run a datacenter-scale model.

VLA-Pro stores task-specific "procedural memories" (basically LoRA adapters) during training and retrieves relevant ones at inference time. The headline number is wild: real-world success rate jumping from 5.8% to 65.0%. That's a 207% relative improvement in simulation.

These VLA approaches feel like the current hype cycle, sort of like how transformers ate computer vision after eating NLP. The question is whether the fundamental assumptions hold. Language models work because text is, basically, a solved representation problem. Robot actions don't have that luxury.

So What

Look, I'm not going to tell you that any of these papers solves robot learning. They don't. The data problem is structural: robots operate in physical reality where mistakes have consequences, and you can't just scrape the internet for more training examples.

But what I see in these six papers is a maturing field that's getting more honest about its constraints. The Phantom team isn't claiming human videos replace robot data forever, they're showing a specific technique that works for specific tasks. The VE2VF people aren't saying vision is useless, they're arguing that sensor fusion beats vision-only for robustness.

This is how real progress happens. Not with revolutionary breakthroughs that change everything overnight (I've seen that movie before, it usually ends with a quiet retraction), but with incremental improvements that stack up over time.

The kids building these systems are doing serious work. Whether it adds up to general-purpose robots in five years or fifteen, I genuinely don't know. But the trajectory is more interesting than it was five years ago, when everyone was just throwing more compute at the problem and hoping for the best.

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

James Chen · 4 hours ago · 5 min