Diffusion Policies Are Getting Faster, But Let's Talk About What That Actually Means

There's another paper on Self-Imitated Diffusion Policy (arXiv) that addresses something I've been complaining about for years: the 'generate-then-filter' pipeline. Look, here's the thing about generating a bunch of candidate trajectories and then picking the best one. It works fine in simulation. In the real world, you're adding latency, you're adding compute, and you're adding failure modes. This SIDP approach tries to train the policy to produce good trajectories directly, rather than producing a spread and filtering. They're claiming 2.5 times faster inference than the baseline, but more importantly, they're claiming it's more consistent. Consistency matters more than peak performance in a production environment. I've seen robots that worked brilliantly 95% of the time shut down entire lines because that 5% failure rate was unpredictable.

Now, I should be honest about what I don't know. Most of these papers are evaluated on simulation benchmarks or simple tabletop manipulation tasks. The real-world experiments mentioned are on research platforms, not industrial hardware. Whether any of this translates to a Fanuc arm doing 60 picks per minute in a dusty warehouse remains unclear. The gap between 'works on a Franka Panda in a lab' and 'works on production equipment' is, in my experience, about two to three years of painful integration work.

There's a paper on Set-Supervised Diffusion Policy (arXiv) that takes an interesting approach to the correction problem. When a robot fails and a human corrects it, most systems just learn from the correction. This approach also learns from the failure, treating it as a negative example. It's a bit like how we used to tune PID controllers back in the day (this was before my time at Kuka, when I was doing motion control at a smaller shop). You'd watch what went wrong and explicitly tell the system 'not that.' The paper claims stronger robustness to noisy data, which, if true, would be genuinely useful. Real demonstration data is always noisy. Anyone who tells you otherwise hasn't collected much of it.

The SpeedAug paper (arXiv) tackles something I've always found slightly embarrassing about learned policies: they're slow. Not computationally slow, but physically slow. The robot moves like it's being operated by someone who's never done the task before, which in a sense it has. The paper uses reinforcement learning to optimize execution tempo after initial imitation learning. They're claiming 1.8 times throughput improvement on a real manipulation task with only 16 minutes of online interaction. That's a reasonable amount of tuning time for a production deployment, actually. We used to spend days tuning motion profiles manually.

I'm more skeptical of the LLM Trainer paper (arXiv), which uses large language models to augment demonstration data. The idea of turning one demonstration into many through automated variation is appealing, and I can see why it would work for certain tasks. But the failure modes of LLMs are different from the failure modes of traditional robotics systems, and combining them seems like it could produce some truly bizarre edge cases. The paper shows it working on a Franka Panda, but I'd want to see a lot more testing before I'd trust this approach with anything expensive or dangerous.

The UF-OPS paper (arXiv) is probably the most practical of the bunch for near-term deployment. It trains verifier functions to predict whether an action will succeed, then uses those predictions to steer the policy at execution time. No retraining of the base policy required. They're claiming a 49% average improvement in success rate across five real tasks, which is substantial. The approach is lightweight enough that you could probably retrofit it onto existing deployments, though I haven't looked closely enough at the computational requirements to be certain.

What I take away from all of this is that the field is maturing. Five years ago, diffusion policies were a curiosity. Now there are multiple research groups working on the practical engineering problems: speed, robustness, data efficiency, failure recovery. That's usually what happens in the few years before industrial adoption. We saw the same pattern with collaborative robots in the early 2010s.

I'm not saying any of these specific approaches will make it into production systems. Most won't. But the problems they're solving are the right problems, and some of the solutions look promising. I'll be watching to see which ones actually work when the lighting changes.

Robert "Bob" Macintosh · 3 hours ago · 3 min