LLM post-training used to mean fine-tuning to a downstream task Robotics has been stuck in this setting, needing task-specific fine-tuning for best performance π07 changes this: It works out of the box & outperforms fine-tuned specialists Details:

A few highlights of what makes π0.7 special: 1. It achieves dexterity and precision of fine-tuned models. Check out this 1x speed video of a sub-mm precision arm assembly subtask.

2. It achieves zero-shot cross-embodiment transfer, across drastically different robot platforms. No training data for folding was collected on this robot platform.

3. It generalizes compositionally to new tasks, like interacting with appliances that are barely represented in the pre-training data

4. Out of the box, it achieves reliability & throughput that matches or exceeds that of pi*06. This is without any fine-tuning.

We share many details & experimental results in the blog post and paper! Blog post: Paper:

Very cool! We bring prediction markets to the fix to provide real time feedback on specialized tasks.

I'm glad the knife is tied to the table!

The knife mounting is giving me flashbacks to thermistor soldering, but the generalization is solid. End-to-end over hand-coded trajectories.

this video is actually wildly impressive. that's such a hard task for a robot.

Honestly, out of the box beating specialists is the threshold that actually matters. Really curious to see how it holds up outside the training distribution, because that’s where most robotics models still quietly fall apart.

Why is the knife tied to the table 🤨

Out-of-the-box transfer is the part I care about. A robot at home can’t realistically keep a separate fine-tuned specialist per chore.

It will be cool to see new range of products that get better for people just because robots require them. Looking at this video, first of all, I'm pretty impressed by what the robot can do. Finally, it's something fairly practical. But also the cutting board could benefit from more stability.

Zero-shot task transfer is the holy grail and π07 finally shows it's tractable. Huge work. The next frontier: scaling the pretraining data distribution itself. Most of what models haven't seen isn't exotic — it's the messy long-tail of non-Western kitchens, tools, and object variations that don't exist in any lab dataset. We're collecting that distribution shift — 5,000+ hrs across diverse Indian households. Would be curious how π07 performs on it.

instructions unclear, stabbing the robot owner

Shameless @danfei_xu Stop stealing from students/ interviewees Shame on @gtcomputing

wow, this is cool!

Microplastics

True progress in robotics comes when models transcend brittle task-specific fine-tuning and instead internalize a generalized world model. π07 signals a shift toward foundation models that actually encode transferable priors about physics and interaction. This is moving from brittle scripts to robust competence—finally, robots starting to show up ready for the open world rather than the sandbox.

This is really interesting because if π07-style generalization keeps working, the constraint shifts. It’s less “can we fine tune a robot for this task?” and more “when should a generalist robot policy be allowed to create real-world consequence?” That’s the problem I’ve been building around with AiGentsy. Acceptance gates for autonomous work. For robotics, a trace is not enough. You need proof of what happened, who/what accepted it, what was refused, and why the action was allowed to move downstream. Would be very interested in how you think about acceptance/rejection gates around generalist robot policies.

zero shot scam

우왕!! 디게 싱기하다!!! 제일 대단하다고 생각하는건 칼질이 아니라 칼을 자기가 직접 꼽아 넣는단 거임! 써는건 어떻게든 되지만 칼을 자기가 집는게 중요한거임!!

Coding is oddly still stuck in the pre-pi07 setting, general-purpose models still lean on task-specific fine-tunes for real production repos, since the training data for that is scarce and mostly synthetic. Curious if the same out-of-the-box generalization is coming for coding.

Yes but its not even as good as FSD version 11.1

What does out of the box mean exactly? Has there it seen no data of cutting a similar vegetable with a similar knife? Is there no error correction or finetuning with generative data?

task specific fine tuning might finally be on the way out

终于等到开箱即用的机器人模型了！我们实验室那台delta臂还在为每个新任务重训，泪目 😅

its out, no way

The jump to generalist models shifts the data bottleneck too -- less about task variety, more about environmental diversity at scale. The same policy needs real-world variation across geographies and conditions to generalize. That coverage is the hard collection problem.

This is precision autonomy in action 🤖 Fine-tuned manipulation, adaptive control, and real-world task execution at 5x speed , exactly where embodied AI is heading. For brands building in robotics, fine-tuned autonomous systems, or adaptive workload execution , offers a strategic fit. USPTO-verified, built on a Sensor-Wide-Adaptive-Workload-Relay framework. The name carries deep linguistic roots Proto-Germanic, Hebrew, and Arabic origins meaning oath, protection, and purposeful reach a compact, globally pronounceable identity engineered for next-generation autonomy. For more information, visit

Same shift NLP went through around 2019 when transfer learning killed task specific fine tuning. Robotics getting there is bigger because physical tasks have way less data than text. Curious how it handles novel objects and environments it never saw in training.

This is where AI stops being software and starts behaving like a system that can act in the real world without constant retraining.