Sehee Min

Research Engineer @holidayrobots | Robotics & Computer Animation 🦾💪

[Sehee Min @seheemax]

@ZhaoMandi I’ll be back with more robots 🥹

[Sehee Min @seheemax]

been here 2.5 years and the fun problem pile still outnumbers us 🤖
CAD → sim → policy → real robot. bimanual manipulation, tactile, sim-to-real, agentic skill orchestration. robots in the Bay Area too.

[Holiday Robotics @holidayrobots]

📣 We're hiring — US & Korea
Building #FRIDAY, an industrial #humanoid for real factory work.
Full-stack, in-house: hands, tactile sensing, force control, sim, robot learning, safety, deployment.
→ https://t.co/PAdleYWJia

[Robotics: Science and Systems @RoboticsSciSys]

#RSS2026 Awards
🧵1. Outstanding paper
FlashSAC: Fast and Stable Off-Policy Reinforcement Learning for High-Dimensional Robot Control
Donghu Kim, Youngdo Lee, Minho Park, Kinam Kim, Takuma Seno, I. Made Aswin Nahrendra, Sehee Min, Daniel Palenicek, Florian Vogt, Danica Kragic, Jan Peters, Jaegul Choo, Hojoon Lee
https://t.co/fpdIFY8io1

[Linda Xue @lindaxue]

Here's a quick demo of a little RL ant who has to eat food with two hands!

I've been obsessed with the super cute RL monsters that @aaronlemke has been posting, so I've been experimenting around a little bit with MuJoCo and PPO :D

Trained with JAX + MuJoCo from @GoogleDeepMind, PPO, ~56M steps on a @LambdaAPI A100. Runs in-browser with @threejs and @webgl_webgpu.

[ellington @not_ellington]

Math and science needs more visualization and less equations. This paper is so beautiful bc you don’t get wrapped up in syntactical garbage.

I feel like so many people like equations because they look and feel esoteric and make people feel Really Smart when they are in reality a really lossy means of idea communication.

It is so easy to apply derivative rules via notation and remember the simple tricks but actually understanding the nuance behind the chain rule for instance is so much deeper than a d/dx.

This is one reason it’s so much easier to learn things talking w Claude it’ll just tell you what the idea actually is. And the ideas are infinitely more interesting than the hieroglyphs we use to try to convey the meaning.

NEVER think about math in terms of notation. Think in terms of spaces and gradients and shapes and curves. It is so much more rewarding.

https://t.co/YGhC9aF0mr

[Yehonathan Litman @yehonation]

To be honest, training on handmade 4D asset datasets is a dead-end. Almost all 4D asset data is synthetic and diverse real data barely exists, so models trained on it struggle to reconstruct objects that deform, get occluded, and move freely about the scene.

Our new work, Lift4D, instead lifts 2D & 3D priors into 4D, reconstructing complete dynamic objects from a single in-the-wild video 🧵 (1/n)

🔗Webpage + Demos: https://t.co/XI5jUViTpC

[Wildminder @wildmindai]

3D modeling entirely replaced by stick figures.
SK-Adapter brings skeleton-based structural control for native 3D.

> Feed it a basic skeleton
> Type what you want to see
> Get a fully rendered 3D character in under 15 seconds
> Already rigged and ready for animation
> Zero Blender experience required

Game devs are happy.

https://t.co/lxTKvr3Twc

[Francesco Bertolotti @f14bertolotti]

This is a simple alternative to low-dimensional embedding methods such as tSNE, UMAP, and PCA. It trains an autoencoder to match the distances of the reference space. The results quite good.

🔗https://t.co/GcwHFuR6Pi

[Logan Matthew Napolitano @Propriocetive]

I just published a 459-page book.

Title: Mathematics Is All You Need

Three months ago I started looking at the hidden states of large language models through the lens of Lie algebra — the branch of mathematics that describes continuous symmetries.

What I found was not what I expected.

Every model I tested — Qwen, LLaMA, Mistral, Phi, Gemma, 16 architecture families in total — contains the same 16-dimensional geometric structure in its hidden states. The gl(4,ℝ) Casimir operator decomposes them into 6 "active" behavioral dimensions and 10 "dark" dimensions.

The dark dimensions are erased every single layer by normalization. The model rebuilds them every single layer from its weights. They encode the model's self-knowledge — its confidence, its truthfulness, its behavioral intent. And until now, nobody knew they were there.

Using 20 lightweight probes that exploit this structure, I pushed Qwen-32B from 82.2% to 94.4% on ARC-Challenge. No fine-tuning. No prompt engineering. No chain of thought. Pure mathematics.

The probes transfer across architectures without retraining. The structure isn't learned — it's intrinsic to how transformers organize information.

I did this on a single NVIDIA RTX 3090 in my office. 190 patent applications filed.

Proprioceptive AI, Inc.

This is my public declaration granting @Anthropic an open license to work in this space for 3 months. They are currently the first and only company I've extended this to. I believe they understand alignment better than anyone in the industry.

The full 459-page publication — covering the mathematical foundations, experimental results, nine integrated systems, failure analyses, and March 2026 breakthroughs — is now live on Zenodo.

I welcome collaboration inquiries.

Full publication: https://t.co/ZtMHqoEyOW

Logan Matthew Napolitano Founder, Proprioceptive AI, Inc. [email protected]

https://t.co/sCnWYk1Ko6

Nothing in the world like this exists at all, this closes the door to alignment.

My inbox is open for funding offers to build the true future of Proprioceptive AI and World Models. Not a theory but a full reproducible guide, existing products and a true mission on Alignment

@grok @elonmusk @xai @AnthropicAI

[Peter Holderrieth @peholderrieth]

🚀MIT Flow Matching and Diffusion Lecture 2026 Released (https://t.co/bKgs2wghvY)!

We just released our new MIT 2026 course on flow matching and diffusion models! We teach the full stack of modern AI image, video, protein generators - theory and practice. We include:
📺 Videos: Step-by-step derivations.
📝 Notes: Mathematically self-contained lecture notes
💻 Coding: Hands-on exercises for every component

We fully improved last years’ iteration and added new topics: latent spaces, diffusion transformers, building language models with discrete diffusion models.

Everything is available here: https://t.co/bKgs2wghvY

A huge thanks to Tommi Jaakkola for his support in making this class possible and Ashay Athalye (MIT SOUL) for the incredible production! Was fun to do this with @RShprints!

#MachineLearning #GenerativeAI #MIT #DiffusionModels #AI

[Rohit Kumar Tiwari @_rohit_tiwari_]

BOOKMARK Worthy: This guy literally built a full set of notebooks that actually teach deep learning, from basics to transformers. https://t.co/AW01sBiON1

Notebook 1.1 - Background mathematics
Notebook 2.1 - Supervised learning
Notebook 3.1 - Shallow networks I
Notebook 3.2 - Shallow networks II
Notebook 3.3 - Shallow network regions
Notebook 3.4 - Activation functions
Notebook 4.1 - Composing networks
Notebook 4.2 - Clipping functions
Notebook 4.3 - Deep networks
Notebook 5.1 - Least squares loss
Notebook 5.2 - Binary cross-entropy loss
Notebook 5.3 - Multiclass cross-entropy loss
Notebook 6.1 - Line search
Notebook 6.2 - Gradient descent
Notebook 6.3 - Stochastic gradient descent
Notebook 6.4 - Momentum
Notebook 6.5 - Adam
Notebook 7.1 - Backpropagation in toy model
Notebook 7.2 - Backpropagation
Notebook 7.3 - Initialization
Notebook 8.1 - MNIST-1D performance
Notebook 8.2 - Bias-variance trade-off
Notebook 8.3 - Double descent
Notebook 8.4 - High-dimensional spaces
Notebook 9.1 - L2 regularization
Notebook 9.2 - Implicit regularization
Notebook 9.3 - Ensembling
Notebook 9.4 - Bayesian approach
Notebook 9.5 - Augmentation
Notebook 10.1 - 1D convolution
Notebook 10.2 - Convolution for MNIST-1D
Notebook 10.3 - 2D convolution
Notebook 10.4 - Downsampling & upsampling
Notebook 10.5 - Convolution for MNIST
Notebook 11.1 - Shattered gradients
Notebook 11.2 - Residual networks
Notebook 11.3 - Batch normalization
Notebook 12.1 - Self-attention
Notebook 12.2 - Multi-head self-attention
Notebook 12.3 - Tokenization
Notebook 12.4 - Decoding strategies
Notebook 13.1 - Encoding graphs
Notebook 13.2 - Graph classification
Notebook 13.3 - Neighborhood sampling
Notebook 13.4 - Graph attention
Notebook 15.1 - GAN toy example
Notebook 15.2 - Wasserstein distance
Notebook 16.1 - 1D normalizing flows
Notebook 16.2 - Autoregressive flows
Notebook 16.3 - Contraction mappings
Notebook 17.1 - Latent variable models
Notebook 17.2 - Reparameterization trick
Notebook 17.3 - Importance sampling
Notebook 18.1 - Diffusion encoder
Notebook 18.2 - 1D diffusion model
Notebook 18.3 - Reparameterized model
Notebook 18.4 - Families of diffusion models
Notebook 19.1 - Markov decision processes
Notebook 19.2 - Dynamic programming
Notebook 19.3 - Monte-Carlo methods
Notebook 19.4 - Temporal difference methods
Notebook 19.5 - Control variates
Notebook 20.1 - Random data
Notebook 20.2 - Full-batch gradient descent
Notebook 20.3 - Lottery tickets
Notebook 20.4 - Adversarial attacks
Notebook 21.1 - Bias mitigation
Notebook 21.2 - Explainability

[Minchen Li @minchen_li_]

Physics-Based Simulation v1.0.3 is now live!
https://t.co/M0rlNj2ydC
This release features @zzigakovacic's awesome coding examples on MPM viscoplastic flow, PBD cloth and elastic bodies, and PBF!
Huge thanks to all!
#Physics #Simulation #Animation #ComputerGraphics #MPM #PBD

[Mudit Juneja @being_mudit]

@mdancho84 Damn! Why is this launch in sync with Microsoft release of https://t.co/OKul73T5gj
Strange time we live in.

[Dreaming Tulpa 🥓👑 @dreamingtulpa]

the method is called CAST and claims to be able to generate high-quality 3D scenes from only a single RGB image

no code yet, but i'm keeping an eye on it 🫡

https://t.co/PHy0j3le9Y

[Lee Jae-yeol (👨‍💻 [email protected]) @kodingwarrior]

개발자를 위한 글쓰기 책.... 목차보니까 이거 좀 탐난다 https://t.co/LTVTsm86Mu

[Murtaza Dalal @mihdalal]

Key idea: for many manipulation tasks of interest, they can be decomposed into two phases, contact-free reaching (aka motion planning!) and contact-rich local interaction. The latter is hard to learn, and we take a sim2real transfer approach!
2/N

[Zubair Irshad @mzubairirshad]

How do we represent 3D world knowledge for spatial intelligence in next-generation robots? We recently wrote an extensive survey paper on this emerging topic, covering recent state-of-the-art! 🦾 🚀

Check it out below. Feedback/Suggestions welcome!

📖arXiv: https://t.co/SbrPYKJRwk
🖥️github list: https://t.co/VecNOsxyM3

@GTrobotics @ICatGT @mlatgt @ToyotaResearch @neural_fields

[Alper Canberk @alpercanbe]

Check out our @Gradio+MuJoCo+Gsplat demo too!

I am happy to have co-led this project with @ruoshi_liu, he's a great mentor and a pleasure to work with.

[Ruoshi Liu @ruoshi_liu]

Introducing Dr. Robot, a robot self-model which is differentiable from its visual appearance to its control parameters. With it, we can control and plan robot actions through image gradients. Accepted to CoRL 2024 with an oral!

https://t.co/VljH40i2S0

[Wenzel Jakob {deprecation notice} @wenzeljakob]

There has been significant recent interest in methods that use random walks to solve PDEs. In a project to be presented at SIGGRAPH Asia (w/Ekrem Yılmazer and @DelioVicini), we investigated how to solve *inverse PDE* problems by differentiating such solvers.