VLA Models Are Getting Faster, But Nobody's Talking About the Real Problem

The Mixture of Horizons approach is more interesting. They're basically running multiple prediction windows in parallel and fusing the outputs. The 99% success rate on LIBERO sounds impressive until you remember that LIBERO is a simulation benchmark. I'd love to see those numbers after a week on a dusty warehouse floor with variable lighting.

Key things these papers are actually solving:

• Threading optimization to reduce end-to-end latency (the agricultural paper got meaningful gains here)
• Decoupling condition monitoring from action generation (the CLAW framework does this for weight thresholds)
• Identifying "transition points" in motion to decide when to re-query the model
• Pruning transformer layers while maintaining performance (DeMaVLA claims this reduces training and inference cost)
• Exploiting temporal redundancy, basically not re-computing when nothing's changed

The REIS paper from arXiv is probably the most honest about what's happening. They observe that "consecutive observations frequently produce identical actions and subgoals." So they built a system that skips unnecessary reasoning. It's not glamorous, but it's practical.

I called my old colleague at Siemens last week to ask what they're seeing in customer deployments. He laughed. Most of their industrial clients are still running deterministic motion planners from the 2010s. The VLA stuff is, in his words, "five years out from being three years out."

That might be pessimistic. The DeMaVLA paper pre-trained on roughly 5,000 hours of real-world dual-arm demonstrations, which is a serious data investment. And their human-in-the-loop correction pipeline (they call it DAgger) addresses one of the fundamental problems: models trained on perfect demonstrations fail on edge cases, so you need to train on failures too.

I'll be honest, I don't know if any of this translates to the kind of reliability industrial customers need. The papers report success rates in the 60-80% range for many tasks, with some hitting 99% on specific benchmarks. That sounds good until you calculate what 80% success means across 10,000 picks per day.

The agricultural manipulation paper is actually doing something useful by targeting low-cost robotic arms. If VLA models can run on cheap hardware handling garlic bulbs and walnuts, that's a market nobody's serving well right now. Whether the threading optimizations hold up outside controlled conditions remains unclear.

What I'm watching for: someone needs to publish long-duration reliability data. Not "we ran 100 trials in a lab" but "we ran 100,000 cycles over six months and here's what broke." Until then, this is all very promising research that I wouldn't bet a production line on. Yet.

Robert "Bob" Macintosh · 1 hour ago · 4 min