Ahmed Agha

Robot Learning PhD Student @northeastern

[Ahmed Agha @aagha314]

Working with Elie is super fun, also you rarely see someone with the mix of experiences that he has in ML and robotics

[Elie Aljalbout @EliJalbout]

We're hiring! The Seattle Robotics Lab (SRL) at NVIDIA is hiring a research scientist to work on robot foundation models. The focus will be on investigating the science behind building VLAs, WAMs, and world models for robotics, and leveraging the findings to propose the next generation of models and training paradigms.

We're seeking a world-class researcher (PhD grad) or senior researcher (3+ years of post-PhD experience) with expertise in training generalist policies. The researcher should aim to question conventional wisdom, push beyond the status quo, and dive into the foundations of general-purpose robot intelligence.

NVIDIA SRL provides a rare environment to conduct open research on fundamental and applied robotics, with the freedom to publish, collaborate broadly, and tackle big scientific questions at scale.

If this sounds like a strong fit, please reach out to us and apply here!

https://t.co/F6XHzOnLDn
https://t.co/extfNzBMDe

[Mathieu @miniapeur]

Welcome!

[gabriel @g_br_l]

when I have to kill the long running session that has all the good info and a constructive back and forth because claude just starts getting context window dementia

[Ryan McWhorter @rmcwhorter99]

give me six hours to chop down a tree and i will spend the first four figuring out how to give to axe to claude

[Ahmed Agha @aagha314]

@CSProfKGD On the one hand that's bad, but also if you're reviewing for robotics+ml+cv conferences you're not gonna have any time to do anything else, i've had to decline aaai's invitation to review this cycle, bc i'm tired of reviewing frankly (and they have 50k submissions 💀)

[Jonathan Gallagher @BeingJonG]

Bro don't tell me the age of research is already over and i'm not even close to being done my PhD

[Ahmed Agha @aagha314]

This post is really good. I think foundation models stuck on some flow-matching head have sucked a lot of oxygen out of robot learning research, maybe we should still explore a bit and not stick to exploiting

[Haojie Huang @HaojieHuang13]

Is Manipulation Policy Learning Not Sexy? An Introduction to the Headaches of Robot Learning https://t.co/5uXJkCCSdY

[Ahmed Agha @aagha314]

@VilleKuosmanen surprised pi0.5 > molmoact2

[Ahmed Agha @aagha314]

I watched Leslie Kaelbling talk about code as a policy a couple of months ago, and the performance was quite impressive. I don't think there's a clean comparison between them and learned policies though, which makes things hard to assess

[Finbarr Taylor @finbarr]

Incredibly cool company. Claude Code for robotics- prompt to policy in 20 mins. Think we're going to see an explosive wave of new robotics companies built on this technology.

[Myk is Walking Backwards @mykola]

me, begging, crying, on my knees: "Please just use plain english, I don't understand what you're saying."

Claude: "The right fix, and the book's lesson applies: the tutorial rotted as a front door because a front door full of claims always rots. So the new root holds only what's timeless — the thesis and three doors — and every claim lives behind it in the thing that goes red when stale."

[Ahmed Agha @aagha314]

that's how it feels tbf

[ily⚡️ @0xIlyy]

Opus 5 cybersecurity requests fallback to Opus 4.8 too 😂

[Tonghe Zhang @TongheZhang01]

Robotics: operation > algorithm.

You will strongly agree with me if you deal with real robot and data at scale. You will strongly disagree if you’re satisfied with overfitting toy demos or do theory and vibe checks in sim.

[Ahmed Agha @aagha314]

amazing

[Susan Zhang @suchenzang]

this moment would not be complete without uncle jurgen coming in like a ray of sunshine rising from the sands of arabia

forever a jurgen stan 🫡

[Bill Gurley @bgurley]

Today in AI.

[Holger Zschaepitz @Schuldensuehner]

Good Morning from #Germany, where a founder boom is defying the economic gloom. A record 3,053 startups were launched in H1 2026, up 88% YoY. 1 in 3 focuses on AI, while VC funding jumped 44% to €5.8bn. Germany’s next growth engine may be emerging.

[Alex Konrad @alexrkonrad]

Damn is there anything these new Chinese models can’t do

[huaijiang @huaijiangzhu]

my spicy theory is that chinese ai labs keep winning because of culture, not talent or resources.

chinese labs still have a deeply hands-on engineering culture. deepseek and kimi are flat organizations where science and engineering are fused: the same people move between algorithms, data, and infra, doing whatever it takes to make the model work.

but sf tech bros have decided that “researcher” is the high-status title while infra is merely support work. every new sf ai startup calls itself a “lab,” every ambitious engineer quietly upgrades their title to “research engineer,” and the infra work is left to whoever failed to escape it.

but at frontier scale, infra determines experiment velocity, and experiment velocity determines research output. at frontier scale, infra **is** research.

[Ahmed Agha @aagha314]

thanks twitter algorithm

[Ahmed Agha @aagha314]

In our experiments

R2-Dreamer: Redundancy-Reduced World Models without Decoders or Augmentation
and
Next Embedding Prediction Makes World Models Stronger

showed better probing performance but were just not as hyped

[Haiyu Wu @HaiyuWu1]

LeWorldModel - what you may know, what you may not realize.

You might know:
Based on the advantages of SIGReg, this is the first latent world model that is insanely simple and cheap to train, achieving performance on par with or better than DINO-WM.

You might not realize:
1. All the tests here are open-loop. It means the model "imagines" the action sequences to reach the goal, then executes them in the real simulator to evaluate the success rate.
2. The violation-of-expectation evaluation results tell us that LeWM learns physics rather than just remembering colors or shapes. However, it is not that sensitive to 3D tasks.
3. Temporal straightening results indicate that LeWM can emerge with a more temporally straightened latent space (this is what I did not realize). This means that LeWM naturally learns a latent space that is easier for planning. The theory is from the paper I posted before (https://t.co/t31pITwRAS).
4. One weakness of LeWM is that it needs action labels, so training a latent world model without action labels is a very important topic.
5. Some configurations and results differ within the paper or between the paper and the code. For example, the two paragraphs describing CEM in the appendix are different. Also, using the default eval seed on Reacher cannot reproduce the success rate reported in the paper. I think we should use multiple seeds for both training and eval in this case because two eval seeds can result in different performance (up to 10% in my experiments).
6. You might have heard from AI agent(s) that, in one action block, all frames use the same action, but it is actually a concatenation of the individual actions from each frame.

It is a good start, but definitely not the end.

[Seth Raphael @magicseth]

A junior dev asked the staff engineer: "How do I know if the AI's code is correct?"
The staff engineer replied: "How do you know if yours is?"
The junior dev was enlightened. Prod went down.