We’re still far from robots that can do anything, but π0 is tackling real-world tasks: laundry folding, takingble bussing and more…

☕️ We’re inching closer to solving the Wozniak test—a robot making coffee in an unknown kitchen.

Check out my full notes ➡️
www.cloderic.com/content/2025...

It follows a two (actually three) phases training process:

🏝️ off-the-shelf VLM training
🤖 VLA Pretraining → Builds a generalist skillset from huge datasets (mix of proprietary + open-source).
🔧 VLA Task-specific Post-training fine-tuning → Sharpens performance, especially for complex tasks.

LLMs proved that generalist models can be surprisingly good across a range of tasks, making AI capabilities easier to demo (and sell). π0, the first model from Physical Intelligence, takes a convincing step toward bringing that same flexibility to robots.

Cool stuff, but I have some notes 😉 They focus heavily on expressiveness metrics without really considering efficiency or utility. Also, the study relies on pre-recorded videos, which could introduce bias from the human actor.

Full breakdown & thoughts here: www.cloderic.com/content/2025...

Also one on the only platform that doesn’t have significant competition with similar offerings. Maybe they one leads to the other…

Just finished the section that’s already online.

It is great content, both succinct and very complete as an intro. I think I’m gonna recommend it to a lot of people!

Congrats to the @hf.co team!

www.cloderic.com/content/2025...

Well yeah, my point is I’m not sure it’s really new (on the defense + AI front).

What seems new-er is more tied to the US politics at the moment and the unchecked (hopefully just for now) that a certain group of people got.

I’m not sure I get what you mean here, AI & Defense industries has always worked closely, for a lot of different reasons including the unique ability of the defense industry to play the long game.

For example, I think there’s room for AI agents that generate code while maintaining dynamic human oversight—something the paper doesn’t fully consider.

Check out my full notes here: www.cloderic.com/content/2025...

Ahh ok, I thought it was a direct way of saying “click subscribe and don’t forget to ring the bell” and I already rang the bell !!

Anyway the rest of the season is gonna be fun if it end up being a take on human annotators.

I’m subscribed :) did I miss something ?

Lol je me disais qu'à l'occasion de la peut-être mort du NFP, on se demandait comment prononcer le prédécesseur :D