
Postdoc @VcaiMpi MPI for Informatics | Virtual Avatars, Human Motion, Physical AI | Prev. PhD @Polytechnique, @Ubisoft @CYENSCoE @la_UPC | Dad 👦🏻👶🏻
Stuttgart, Germany
Joined June 2020
- Tweets331
- Following608
- Followers362
- Likes631
Pinned Tweet
Imagine X-ray vision from video alone 🩻
Releasing SOMA v2! #ECCV2026
• SKIM v2: cleaned dataset & rigs, fixed hands
• New pretrained models
• 58% lower marker error than skinning (LBS)
• 0.06% muscle - skin collisions
Enjoy the video! Code + data 👇
github.com/edualvarado/SOMA
Body models capture skin, not what drives it. Muscles stay invisible - recoverable only via MRI/CT or costly simulation, neither of which scales to full-body motion.
Can muscle deformation be recovered directly from vision/how skin moves? We present SOMA at #ECCV2026!
🧵
7
605
PPO, part 1: acting. How 14 numbers out of a network become a duck moving on @pollenrobotics' Microduck.
Left: the trained policy, mean only.
Right: the same policy with training noise, σ 0 → 1.
Post + interactive demo: edualvarado.com/miniblog#ppo…
2
106
61 numbers in → 14 means → + Gaussian noise in training (σ starts at 1.0) → target = home + action, never clipped → XL330 servos push, at most 0.64 Nm, 15–30 ms late.
On the real duck, the noise is gone: only the mean is used.
1
31
The demo runs the Microduck in MuJoCo in your browser, with Pollen's policy as plain JavaScript: edualvarado.com/miniblog/ppo…
Part 0, the batch: edualvarado.com/miniblog#ppo…
49
Part 0, the batch: 4,096 ducks × 24 steps = 98,304 samples → 20 gradient steps.
The series follows one iteration in the order it runs:
0 Batch · 1 Act · 2 Reward · 3 Critic · 4 Advantage · 5 Clip · 6 Passes
Every number comes from the robot's real training config.
1
42
Demo on its own: edualvarado.com/miniblog/ppo…
The survey: arxiv.org/abs/2203.04735
Microduck model, training code + the walk in the video: @pollenrobotics (CC BY-NC-SA 4.0). Trained with mjlab + rsl_rl.
47
After revisiting the survey on RL that I co-authored eons ago with @ArielKwiat, I will be building some short demos of the core ideas.
First up: PPO, taken apart one training stage per post, on @pollenrobotics' Microduck.
Post + interactive 3D demo: edualvarado.com/miniblog#ppo…
1
2
204
Eduardo Alvarado retweeted
Replying to @ChongZzZhang
Exactly. My concern isn’t with the authors or how quickly they built on existing work. It’s with the surrounding hype that research can now be done in 2–3 days and that peer review no longer matters. We should celebrate advances without creating misleading expectations for younger researchers.
Insane work!
We found a way to scale humanoid control that reduces GPU-hours by more than 1000×.
Across 129,514 motions, SuperMimic surpasses the state of the art (128 GPUs, 7 days) with 1 GPU in 3 hours.
Project: supermimic.github.io
Demo: supermimic.github.io/demo/
1
86
Before LLMs, we were limited by time, knowledge, and taste. A lack of the first two is no longer an excuse not to build what you have in mind.
That’s a hard pill to swallow for some - knowing their imagination was the actual barrier all along.
1
78
Retro-futurism is the best aesthetic ever created. Change my mind.
Here's an in-depth look at the new @Tesla Roadster's UI. Love the retro-futuristic design aesthetic.
77
I have always thought the key for manipulation scaling is huge amounts or extremely rich data, far richer from what we have today. The softness of our skin plus the precise sense of pressure and friction.
The question is, how we simulate/record these at scale.
Touch is essential to human manipulation. Yet, we have so far not had a tactile sensor that can do what human skin can: sense texture, firmness, vibration at a viable form factor. At Rhoda, we have built a new sensor in-house that can do all of these things, while being much much smaller than commercially available ones!
66
This makes all the sense of the world and will become a standard soon.
Introducing Code Video Model 🎬
Amazing work led by @raywzy1 and the team!
Agentic coding as the engine for controllable video generation.
Scene layout, motion, timing, camera trajectories, and physical dynamics, all in editable code.
A coding agent turns prompts into executable programs that render proxy videos. Our training-free early-step conditioning lets a pretrained video model follow the proxy’s structure while adopting a reference image’s appearance.
From physics-grounded motion and bullet time to 3D/4D reconstruction and robotics simulation, it all starts with code.
Watch the demo and explore more visual results 👇
Project page: code-video-model.github.io/
Paper: code-video-model.github.io/t…
Code: github.com/code-video-model/…
Data: huggingface.co/datasets/code…
Youtube: youtube.com/watch?v=dTtiVkSR…
43
This is so cool. It’s so refreshing the see companies stepping out from the common trends and see such novel ideas.
After six years in stealth, Atomic Machines is out. Our mission: on-demand, universal command of matter. First beachhead: the Matter Compiler, an AI-native manufacturing system that builds working micro-machines from code alone. No per-product tooling. No process development. Different code, different machine.
82
I would never recommend having kids to solve loneliness. Children aren’t a cure for adult isolation or a way to fill your social void. There are better, less selfish, more meaningful reasons to become a parent.
Will you feel less lonely? Sure, but don’t let that drive you.
Replying to @levelsio
I can also recommend having kids! Nobody is lonely in the presence of a four year-old who just wants to talk, play, and learn with you.
1
3
146














