Generative models (diffusion/flow) are taking over robotics 🤖. But do we really need to model the full action distribution to control a robot? We suspected the success of Generative Control Policies (GCPs) might be "Much Ado About Noising." We rigorously tested the myths. 🧵👇

Common wisdom says GCPs beat Regression because of 1️⃣ Multi-modality handling 🔀 2️⃣ Better pixel-based control 🖼️ 3️⃣ Expressivity 🧠 💡Finding1: We debunked this. When controlling for architecture, regression match flow on almost all tasks, except for a few high-precision tasks.

To understand that gap, we proposed a new taxonomy for GCPs. They aren't just one thing; they are 3 components: 📊 Distribution Learning 🎲 Stochasticity Injection 🔄 Supervised Iterative Compute

💡 Finding 2: It turns out the magic ingredient isn't "Generative Modeling"—it's 🔄 Iterative Correction + 🎲 Noise. We built the Minimal Iterative Policy (MIP): ✅ A simple 2-step regression policy ✅ No complex distribution matching ✅ Matches Flow/Diffusion performance

Apart from single-task settings, our simple recipe (MIP) even works for finetuning VLAs!

💡 Finding 3: Why does Iteration + Noise work? It’s not just "expressivity." It’s Manifold Adherence. Iterative compute ensures that even when the robot sees weird observations (OOD), the output action remains on the "manifold" of physically plausible movements.

Start treating your policy architecture as an iterative solver, not just a distribution fitter ;) 📄 Paper: 💻 Code: 🌐 Website:

If you see @max_simchowitz @GuanyaShi @nmboffi @guannanqu at NeurIPS, feel free to chat with them about our work! Work led by me w/ Giri Anantharaman, Nai-Chieh Huang, Claire Jin, Daniel Pfrommer, Chenyang Yuan, Frank Permenter, @guannanqu , @nmboffi, @GuanyaShi, @max_simchowitz

Insightful work! Thanks for sharing. A quick question, how will MIP behave when data is a.1 (and I agree that most cases we have b.1)?

It depends which loss norm you use, L1 will predict one mode (medium value) while L2 will predict the mean.

So you are saying that if data is a.1, better still using flow matching policy?

I am not sure if those benchmarks in the paper (robomimic, libero,eg) are actually multimodal enough. If both regression and diffusion get near 100% SR, that means the benchmark is not a good one to begin with. Maybe try d3il?

We have pusht and kitchen which is quite multimodal. Do you think d3il is multimodal? Happy to give it a shot ;)

I think these guys have already given a similar theoretical analysis in their paper:

great work! have you investigated matching training FLOPs between these methods? MIP has two forward passes in the training objective

That’s a great question. Computation are aligned since we randomly sample one step in MIP to compute loss. So the function evaluation number stay the same across training.

Great work! Really interesting paper. I'm curious about what you think about recent non-iterative generative policies (C1+C2) like and These methods are basically regression but with additional mechanisms that encourage better use of the noise space. It seems that either C1+C2 and C2+C3 can work well, but I wonder about the trade offs.

In the non-iterative methods you mentioned, their evaluated task (robomimic, kitchen) is where we find regression is on par with flow with correct architecture. In our paper we use shortcut models as C1+C2 example, whose performance is worse than MIP/flow on most challenging task

Haven’t entirely finished reading because it’s a long one but love this type of work!! Good old fashioned science that truly tries to understand things instead following the hype

Great analysis Chaoyi!

Generative Control Policies are getting hype, but testing the assumptions is exactly what the field needs.

Nice work! Do you explore how different models generalize to new objects/envirments? I feel that might be where diffusion/flow models can shine

We have multi-task training with LEBERO. For zero-shot experiment, I guess we need to pertrain a model and this paper doesn’t answer that. I think that would be an interesting future work!

I had Grok explain this to me and why what you reported is important. Then in a way a child could understand. Thought you'd find it entertaining, because it made it very clear to me. Thanks for your work!

great work. recently I was trying to use the mean flow actor type in it.But it seems like there are some problem in your meanflow implementation.The training process is break out in 150k steps.may I discuss with you if you are thinking about meanflow policy in robotics control

Check DM

Shameless @danfei_xu stop stealing from students/ interviewees - Shame on @ICatGT