1/ We just released π0.7 — a steerable generalist robot model with emergent capabilities. I want to share a bit of the backstory, because π0.7 taught me something surprising about where robot learning is heading. A thread on bittersweet lessons 🧵

2/ About a year ago, I started a research project on world models for robotics. My bet was that world models would be the key unlock — that they'd dramatically outperform standard VLA approaches for task generalization. And for a while, the results backed that up. We were getting impressive compositional generalization: robots following unseen instructions, performing tasks never in the training data, transferring from other robots and human videos.

3/ But something funny happened. Our VLA baseline kept getting stronger as we collected more data. Day by day, quietly improving. Until one day it was also showing very promising signs of compositional generalization — and it was much simpler. So what do you do when your baseline eats your research thesis? You write a paper to understand why your baseline got so strong :)

4/ That paper became π0.7, and the lessons I took away have made me much more optimistic about the future of robot learning. Here's what surprised me:

5/ I thought task generalization was hard. Not completely true. I thought learning from heterogeneous data was hard. Not true. I thought cross-embodiment transfer on dexterous tasks was hard. Not true either. Compositional generalization — recombining knowledge and skills to solve new tasks — emerges from training on large and diverse datasets. Even in the action modality.

6/ But it's not just about scaling data. In the paper we show that conditioning and context matter. You have to set up the ML problem correctly. π0.7's key insight: diverse data needs diverse context. Metadata about speed, quality, and visual subgoals disambiguate the training signal. Without that context, heterogeneous data is noisy and contradictory. With it, even suboptimal rollouts and failures become useful training signal.

7/ One result that surprised us: cross-embodiment transfer on dexterous tasks. The convention in the field is to evaluate transfer on pick-and-place. We wanted to try something harder — shirt folding. All our folding data was collected on small, lightweight bimanual robots. We tested π0.7 zero-shot on heavy industrial UR5e arms with imprecise parallel jaw grippers. No laundry folding data for that robot at all. It folded t-shirts consistently — and applied a completely different folding strategy, adapting to the different robot affordances on its own.

8/ Another result I keep thinking about: teaching the robot to use a novel air fryer. When I ordered the first air fryer from Amazon to test this, I had little hope it would actually work. Zero-shot, the robot makes a reasonable attempt but doesn't finish. But if you walk it through step-by-step with verbal instructions — like coaching a person — it handles the appliance effectively. Then you distill those coaching sessions into a high-level policy. The robot learns the task from verbal instructions alone, no additional teleoperation. Prompt engineering, but for robots.

9/ We traced back WHERE the model learned air fryer skills. After a lot of searching through our dataset, we found: a couple of episodes of a static single-arm robot closing an air fryer in someone's home, and a few DROID episodes of a Franka arm opening one. These look nothing like what our mobile robot actually does. The model composed fragments across embodiments, environments, and strategies — much like how an LLM composes concepts from its training data. This is the kind of generalization people have been hoping for in robotics, and it was genuinely unexpected to see it working.

10/ Goal conditioning turned out to be more powerful than I expected. The goal-conditioned policy already knows a lot about achieving goals. A lightweight world model generates visual subgoals that guide the robot and implicitly specify what NOT to do (e.g., don't move your arm around while holding the air fryer tray). Even when subgoals had some mismatch with reality, the policy tried remarkably hard to achieve each one. That robustness wasn't designed — it emerged. I think there's a lot of room to run with goal-conditioned RL from here.

11/ One lesson I keep re-learning: as the model becomes more capable, evaluation becomes the bottleneck. Stochasticity in the environment, the model, and human evaluators all compound. Generalization becomes harder to define and measure — when your dataset is large, it's not easy to figure out what's truly "in" vs "out" of distribution. Eval is a real unsolved problem for the field.

12/ The bittersweet part: I came in with a strong research bet, and reality had a different answer. The bitter lesson strikes again — data had the highest leverage, not the clever method. But here's what's sweet: the results are more exciting than my original thesis. Compositional generalization emerging naturally from data, working across embodiments and tasks, steerable through language and goals. That's a future worth being optimistic about.

13/ This is part of why I love research. You go in with a thesis, nature humbles you, and what you find is more interesting than what you expected. That said, we're far from converging on the right approach — I still see a lot of value in new methods and diversity of research bets in the community. Research bets compound in unexpected ways, even when the answer comes from somewhere you didn't plan. Paper: Blog:

Nice release! If the conclusion is that VLAs outperformed a world model backbone primarily due to data, why can't that same data be leveraged for pre-training/mid-training the world model? The core diff is how the video-action distribution is modeled not the data itself

Amazing work, congrats!!!!

Congrats, Lucy! Excited to read your work.

This is the content I come to X for - amazing work, thanks for sharing. I have to say I walk away excited by this work... but thinking pi0.7 is closer to GPT-2 then it is to GPT3.5.

congrats lucy!! really enjoy reading this

Love these tweets and very impressive work!

Can't wait for more air fryer-operating robots! Very cool!

@chelseabfinn Thanks for sharing the insights Lucy!

Congrats, Lucy!!

CONGRATULATIONS!

lucy, awesome thread. thx!

This is pretty fun, gonna try this out with my model team 😄

Thank you so much for your research and contribution to this field. It really is inspiring. I am an aspiring phd candidate, with interest in VLAs and everything about it. Seeing how fast the field is evolving makes me excited for the future and to learn about it as well.

@richardxp888 Amazing work and analysis!

failures and suboptimal rollouts definitely are useful signals congrats to the team

Amazing work and insights

@chelseabfinn Let’s go Lucy!!

Very cool! World-model assisted VLAs :)

Love the shift to context, curious how you practically curate diverse data?

I'm very curious to know: in your π0.7, during inference, do the high-level subtask prediction and the VLM in the VLA backbone use two pre-trained models?

Nice

VLMs as a backbone are Counterintuitive compared to WMs, but love to see more progress from the team.

Lucy姐的pipeline一如既往的好，点赞🫡

I get the visual subgoal as the high level signal, but that kinda excludes proprio and force/torque from the high-level loop. For contact-rich tasks, do you see non-visual sensors ever entering the sub-goal generator for π?

amazing work

Thanks for the paper, very easy to follow.

lfg

AWESOME

Really great read. Data being the biggest driver for improvement is a pattern we’re seeing across many fields. Interesting that as data scales and becomes more accessible evaluation becomes the bottleneck. How do you think evaluations should evolve to actually capture emergent capabilities?

@svlevine