What Is π0 (Pi-Zero)?

π0 is introduced in the paper π0: A Vision-Language-Action Flow Model for General Robot Control (arXiv 2410.24164) and an official blog post, both from Physical Intelligence. The authors argue that generalist robot policies, which they also call robot foundation models, can address robot learning’s obstacles in data, generalization and robustness. π0 is their prototype for controlling different robots from one model.

π0 is third-party research, and Firsthand is not part of it, endorsed by it or affiliated with its authors. Firsthand makes no claim that its data trained or was used to evaluate π0, and no performance claim against it. In the paper, the training data is robot data plus open-source robot datasets, and egocentric human video is not described as part of the mix. What Firsthand states is custom egocentric collection for VLA-style training: synchronized egocentric video, a paired task instruction and the action that carried it out (3D hand pose, gripper state or a robot action log), hardware-triggered onto one clock, to a written spec, on demand.

The model is one transformer with two sets of weights. Images and language go to a backbone initialized from PaliGemma, a 3 billion parameter vision-language model. Robot state and actions go to an action expert of about 300 million parameters, for 3.3 billion in total.

It predicts a chunk of future actions, 50 steps in the paper’s tasks, by integrating a learned vector field from random noise in 10 steps. The authors state this lets it control robots at up to 50 Hz. Inputs are two or three RGB images per robot, a language command and the robot’s proprioceptive state.

The pre-training mixture combines the authors’ own dexterous manipulation data, collected on 7 robot configurations across 68 tasks, with open-source robot data: the Open X-Embodiment dataset, Bridge v2 and DROID. The authors state it covers over 10,000 hours of robot data, and that open-source data is 9.1% of the mixture by timesteps. Language labels combine task names with segment annotations, typically about 2 seconds long.

Post-training then fine-tunes the model on smaller, higher-quality task data. The authors state the simplest tasks needed about 5 hours and the most complex 100 or more.

They evaluate out of the box, on language following, on fine-tuning to new tasks and on multi-stage tasks such as laundry folding, table bussing and box assembly. Scores are normalized, averaged over 10 trials per task, with partial credit.

They report that π0 outperformed the baselines they compared, including OpenVLA and Octo trained on the same mixture. These are the authors’ results on their own robots and tasks, and Firsthand draws no comparison from them.

The authors state that how pre-training data should be composed and weighted remains open, that not all evaluated tasks work reliably, and that it is unclear how much and what kind of data is needed for near-perfect performance. They also leave open whether the approach extends to domains such as driving, navigation and legged locomotion.

The official openpi repository holds open-source models and packages, including a π0 base checkpoint pre-trained on 10k+ hours of robot data, with fine-tuning examples. Its README says π0 was developed for the authors’ own robots and may or may not work on others. The repository lists the Apache License 2.0; the licence of the checkpoints is Not stated in the sources read.

The README says fine-tuning data is converted to a LeRobot dataset, a format Firsthand also lists for delivery. Firsthand makes no claim that its data fine-tunes π0.

Write the target skills, environments and signals into a skill spec and confirm them on a Spec Pilot, a small fast first batch. This is on-demand custom collection, not a catalog download.

Where coverage is deep, the median from signed spec to first validated batch is 48 hours; where it is thin, three to four weeks. Billing is per validated hour.

What is π0?

A vision-language-action flow model from Physical Intelligence that adds a flow-matching action expert to a pretrained vision-language model to control robots with action chunks.

Is π0 the same as flow matching?

No. Flow matching is the training objective for the action output. π0 is the full model: a vision-language backbone plus an action expert that uses flow matching.

What data was π0 trained on?

Over 10,000 hours of the authors’ robot data across 7 robot configurations and 68 tasks, plus open-source robot data such as Open X-Embodiment, Bridge v2 and DROID.

Was π0 trained on egocentric human video?

The paper describes robot data and open-source robot datasets. Egocentric human video is not described as part of its training mix in the sources read.

Is π0 open source?

The official openpi repository provides π0 base checkpoints and fine-tuning examples. The repository lists the Apache License 2.0; the checkpoint licence is Not stated in the sources read.

Is Firsthand data part of π0?

No. Firsthand is not part of it, endorsed by it or affiliated with its authors, and makes no claim that its data trained or evaluated π0.