🤖🔄🌎 | Senior undergrad, Yao class @Tsinghua_Uni | RA @uwcse | RS @wuji_global | Prev. RA @mldcmu.
Seattle, WA
Joined August 2024
- Tweets339
- Following1.7K
- Followers1.7K
- Likes1.3K
Siqiao Huang retweeted
Wrist cameras give robots close-up views, but only where the hand happens to be.
What if the policy could decide where to look?
Introducing FoveaVLA: a VLA that replaces the wrist camera with learned foveation actions.
FoveaVLA predicts where it will need to look next, then crops that region from the full-resolution external image. The full view gives context; the foveal crop gives close-up detail.
With one external camera and no wrist camera, FoveaVLA reaches 70.8% success on 48 real-world trials, vs. 52.1% for π0.5 with a wrist camera.
• No additional demonstrations
• No camera calibration
• No wrist camera
Paper, code, and models are all open 🧵
Siqiao Huang retweeted
Hey, maybe you are looking for our STARFlow-V (starflow-v.github.io/starflo…) 🤔😲 and it is not diffusion.
Most “autoregressive video” is not autoregressive.
Nobody pretrains BERT, distills a GPT-like student, and calls that native. Video does the equivalent every day: train a bidirectional short-clip DiT, then add a causal mask and a few-step sampler to produce a faster version.
I wrote a short blog on what native would actually require:
waynejin0918.github.io/how-f…
( Thanks to the TML blog for providing stylistic references, and to grok4.6 for helping me polish my content. )
Siqiao Huang retweeted
Can we get robots to think before they act? In Latent Memory Palace, we achieve this through a classic principle: variational inference. (1/8)
Siqiao Huang retweeted
RL can't teach an LLM to solve problems it never solves. Robot policies suffer from exactly the same limitation.
The fix turned out to be almost embarrassingly simple: train the policy with diffusion noise during pre-training.
That's it.
The policy covers a much wider action distribution, RL finally has somewhere to search, and we fine-tune VLAs on real robots in under an hour.
Introducing TMRL.
🧵(1/9)
Siqiao Huang retweeted
Potentially obvious finding: naively taking a policy trained on real-world data and fine-tuning it with RL in simulation can produce quite dangerous behavior…
weirdlabuw.github.io/score/ 🧵👇
Siqiao Huang retweeted
Policies trained on real robot data via imitation can be surprisingly capable. But for domains like dexterous manipulation, they are often not quite good enough: they move slowly, miss grasps, make unreliable contact, and fail under small perturbations.
Can we improve them without any additional data collection on the real robot?
In SCORE, we show that we can improve real-world diffusion/flow policies cheaply by using simulation to simply learn how to steer them on deployment. This leads to large gains in real-world success and speed across a variety of tasks, without requiring additional real-world experience:
weirdlabuw.github.io/score/
🧵 (1/10)
Siqiao Huang retweeted
As a PhD student, I was told not to work on deep RL - too full of hacks and alchemy" But after a year or two of working in this area, I’ve come to (deeply?) appreciate all of the thoughtful research that’s gone into understanding what/why things work and how to make them better.
My lab (joint with @abhishekunique7) took what we’ve learned by reading this body of literature to answer the question: what are the actual best-practices for finetuning a diffusion/flow/generative robot policy (for now, in sim)? Under one set of constraints - ample compute but limited time on your robot - @servo97 paper gives a pretty compelling answer.
Interaction with the real world is the major bottleneck in robot learning. So what would robot RL look like if we didn’t need to limit compute per interaction? Our latest work, Off-Policy Generative Policy Optimization (OGPO, accepted to ICML26) embarks on answering this question (spoiler alert: when done correctly, it helps massively!).
🧵(1/N)
Siqiao Huang retweeted
Super excited to finally announce our new paper “Double Preconditioning (DoPr): Optimization for Test-Time Performance, not Validation Loss”
Tl;dr: from LLMs to robotics, on-policy deployment causes a mismatch between validation loss over the training distribution and “downstream performance”.
We propose Double Preconditioning (DoPr), which drops in an activation-based preconditioner into your favorite gradient-wise whitener/normalizer, such as AdamW or Muon, to help close this gap.
1/12 🧵
Siqiao Huang retweeted
Here’s a pretty weird and surprising result - retrieval-augmented generation works unreasonably well for robot learning – but only when parameterized using difference vectors!
We introduce Difference-Aware Retrieval Policies for Imitation Learning (DARP), a simple, semi-parametric RAG architecture for imitation learning that achieves gains of up to 200% over standard behavior cloning. No additional assumptions beyond BC, just a little architecture switch! The theory backing it up is pretty cool too and it works on real robots! :)
Play with our website to understand better: weirdlabuw.github.io/darp-si…
🧵(1/7)
Siqiao Huang retweeted
Congratulations to Siqiao ( @KnightNemo_ ) for releasing his NanoWorldModels repo.
World model research can be notoriously tricky, given the numerous small design decisions that can make an excellent idea appear to fail completely. Siqiao's repo will help all researchers avoid these pitfalls; a huge service to the community!
In the last couple of months, we have witnessed significant advances in Industry-scale World Models. Yet, for the broader community, the gap between reading about these models and deploying them remains disappointingly wide.
Today we're releasing Nano World Models: a minimalist, batteries-included repo for advancing world model science.
🧵 (1/9)
Siqiao Huang retweeted
Punchline: distill world models from simulation to enable fast, stable real-world robot adaptation.
Simulation is nearly always wrong. But in Simulation Distillation, we ask a simple question:
How do we perform simulation pretraining such that real-world adaptation becomes trivially easy?
sim-dist.github.io
Let's take a closer look (1/n)
Siqiao Huang retweeted
Unlocking test-time scaling and search in generative models has been a fundamental obstacle, and is increasingly important as these methods are deployed in science, engineering, and robotics.
It was a pleasure to work with @peholderrieth on a paradigm that finally makes this capability genuinely possible, and lets you trade off training time vs inference time according to your needs. Check it out!
We release Diamond Maps💎 unlocking accurate and efficient guidance for diffusion models. Our experiments show that our methods scale incredibly well. Excited to see what people will build with this!
Accurate guidance has been a notoriously hard problem, but in this work, we’re bringing TWO (!) solutions to the table. The recipe for success:
1️⃣ Speed: Use distilled models (flow maps, mean flows, consistency models).
2️⃣ Exploration: Inject stochasticity to properly explore your search space.
Because this fundamentally improves anything using flow matching and diffusion, we see a lot of potential for applications across audio, robotics, molecules, and beyond.
Paper: arxiv.org/abs/2602.05993
Code: github.com/PeterHolderrieth/…
Huge thanks to an amazing team: Douglas Chen, @LucaEyring, @ishin_shah, Giri Anantharaman, @electronickale, @zeynepakata, Tommi Jaakkola, @nmboffi, and @max_simchowitz. It was awesome bringing this to life together!
Siqiao Huang retweeted
⚠️Public Service Announcement for ICLR folks headed to Rio: Rio is beautiful, but getting a travel adapter in Brazil is a pain in the ass, and they don't use similar outlets to other major regions (US, UK, Europe, China). Buy a Brazil-specific adapter before traveling. Enjoy :)
Siqiao Huang retweeted
Introducing vla-scratch: a modular, performant and efficient stack for VLAs. github.com/EGalahad/vla-scra…
I started it because existing codebases were either slow, or hard to extend for co-training with clean data abstractions. This repo is a ground-up attempt to address both.
Siqiao Huang retweeted
VLA models are booming—but one fundamental question is yet seldom answered:
How does the choice of the base VLM affect VLA performance⁉️
We run a large-scale systematic study to answer this, in collaboration with @Alibaba_Qwen
cladernyjorn.github.io/VLM4V…
Siqiao Huang retweeted
• Edge-Aware Safety: Novel volumetric edge penalization prevents slipping on terrain edges.
• Extremely dynamic: Robust traversal at up to 2.5 m/s!
• Open-sourced: All codes for training, deployment are open-sourced!
👇 Check out the project page!
project-instinct.github.io/h…