Minghao Fu

Ph.D. Student @UCSanDiego | World Model | Representation Learning |

[Minghao Fu @Minghao__Fu]

JEPAs can 𝐧𝐨𝐰 jointly imagine what happens and what to do next!

Meet 𝐋𝐞𝐖𝐀𝐌: A JEPA for the real world with

32.3× faster planning, 19 ms per plan
Improves success 28.6% → 89.7% on contact-rich manipulations

But also ... 🫶 a WAM for everyone with

17M · 1 GPU · 1 hyperparameter
Works on a real robot, not just a simulation !!

👉 https://t.co/CSSduWapQ7

[Chen Tessler @ChenTessler]

Coding agents can build an app. What if they could give it a body?

Meet Metatron: a versatile System 1 controller for humanoids. A goal-based interface connects System 2 planners, including coding agents, to whole-body movement.

A few examples below. 1/5

[Minghao Fu @Minghao__Fu]

@leothecurious Both are basically goal-conditioned policies, whether the goal is specified through a image representation or a language instruction. If you are interested in a hybrid approach, this paper might be relevant: https://t.co/OuKLTFdYW6

[gokaygokay @gokayfem]

World to World

I gave Claude Opus 5.5 a Gaussian splat of a garden and asked it to rebuild the scene in Blender

Procedural objects, Geometry Nodes, Shader Nodes, then captured textures mapped onto selected surfaces

[Minghao Fu @Minghao__Fu]

True lol, can’t stop ppl from naming their work whatever they want. And tbh, a lot of ppl pumping out AI slop couldn’t care less about their reputation

[davinci @leothecurious]

in an interesting followup...

seems like the Latent evolving World Action Model paper from baidu, initially submitted on sep. 23, only added the "LeWAM" label to the paper title on oct. 10, literally 2 days after a couple of LeWAM papers got submitted (oct. 8), one of which is a followup to @randall_balestr's line of work tracing back to LeJEPA and LeWorldModel, and the other is parallel work from nvidia (which arguably popularized the WAM acronym) building on LeWM in a different direction.

[Minghao Fu @Minghao__Fu]

https://t.co/FhyaJtbX5R is the only end-to-end JEPA with public code, checkpoints, and reproducible results. Led by @randall_balestr as a follow-up to LeWorldModel, the project began quite some time ago.

Personally, I agree others are ai slop or adopt the JEPA label without following its core formulation. Even a quick read raises questions

[Minghao Fu @Minghao__Fu]

@qiayuanliao Super cool

[Minghao Fu @Minghao__Fu]

@DavidSunok47 I don’t want to be too mean, but I spent time reading your paper, and it honestly reads like AI-generated text, with so many self-invented terms and unnecessary jargon. Also, I’ve never heard anyone in RL refer to a model-based policy as “search-free planning.”

[Minghao Fu @Minghao__Fu]

What exactly is the difference between a goal-conditioned policy and “search-free planning with a world model”? Am I missing some new terms for policies 😅?

[Junhan Sun @DavidSunok47]

It’s been 3+ months since INTACT dropped — and it’s already been cited 10+ times 🚀🔥

Huge congrats to HIJP 👏 Really glad to see people learning INTACT, using INTACT, and going beyond INTACT ✨
Real innovation grounded in JEPA theory is rare — big shoutout to the JEPAs 💪

INTACT is still one of the SOTA works on Direct 🎯📈
So… is it time for me to bring you all “some” surprises too? 👀Time for another real shockwave ⚡️

[Minghao Fu @Minghao__Fu]

@DavidSunok47 I agree that ur method has its own contributions and is not just GCBC. But my question remains: how is it different from a goal-conditioned policy??? Is there something wrong with calling it that?

[Minghao Fu @Minghao__Fu]

@DavidSunok47 Different how? input zt zg, output actions, so the difference is just feature engineering?

[Dream Labs @dreamlabs_ai]

Today, Dream Labs is coming out of stealth with Dream-1, a robot foundation model that controls bimanual robots it has never seen, on tasks it was never shown, in places it has never been, from $3/hour.

Dream API is live for everyone, today. 🧵

[Minghao Fu @Minghao__Fu]

@_vztu thats true

[Liang Zheng @LiangZheng_06]

Representation conditioning is better than representation alignment.

The two recent papers reported similar observations. Time for some new trials?

RECON: https://t.co/Zc5HEnU4BN
Dino forcing: https://t.co/yBW8IgNy4f

[Minghao Fu @Minghao__Fu]

@Selen7005717917 Love this proof of concept. A similar work with joint acwm+policy+mpc https://t.co/BFHGOdb5Wr but not dig into so much how w a benefit each other

[Minghao Fu @Minghao__Fu]

@Jaeyeon_Kim_0 @jungyi3578 interesting

[Minghao Fu @Minghao__Fu]

@QuanVng @physical_int Aha, maybe I was just too impressed by the demo. Curious what you think this is still missing for a GPT-3 moment in robotics?

[Minghao Fu @Minghao__Fu]

@nikitabier lol

[Minghao Fu @Minghao__Fu]

@LucasJinReal We haven’t included yet, but if ur interested in bimanual manipulation, take a look at our Transport results. RoboTwin is on our roadmap!

[Minghao Fu @Minghao__Fu]

JEPAs can 𝐧𝐨𝐰 jointly imagine what happens and what to do next!

Meet 𝐋𝐞𝐖𝐀𝐌: A JEPA for the real world with

32.3× faster planning, 19 ms per plan
Improves success 28.6% → 89.7% on contact-rich manipulations

But also ... 🫶 a WAM for everyone with

17M · 1 GPU · 1 hyperparameter
Works on a real robot, not just a simulation !!

👉 https://t.co/CSSduWapQ7

[Minghao Fu @Minghao__Fu]

@TX_Leo_Wang Thanks, Zhi! 🫶

[Minghao Fu @Minghao__Fu]

@ylecun @Helios_Hua @garridoq_ @laurentnajman TLDR:

[Minghao Fu @Minghao__Fu]

@ylecun @Helios_Hua @garridoq_ @laurentnajman Learned a lot from both works! There is another early work that also studies invariant vs. variant representations in dynamics with identifiability (accepted by NeurIPS this year), focus on what the ideal representation for planning is:
https://t.co/F6FZJsanLW