Skip to content
Solutions · Physical AI & Robot Data

Pi0 Fine-Tuning Data for openpi Robot Policies

Pi0 fine-tuning data for openpi: teleoperation episodes for post-training π0 (pi0) from Physical Intelligence and π0.5, recorded on bimanual arms and the Unitree G1, plus UMI handheld demonstrations. Each episode pairs camera views, joint states and commanded actions on one clock, delivered in LeRobot format, ready-made or custom.

Book a Pi0 Data Strategy Session Free 30-minute call · mutual NDA included
openpiEpisodes for the π0, π0-FAST and π0.5 base checkpoints in openpi.
Multi-viewWrist and scene views per episode, the multi-image input π0 reads.
LeRobotThe format openpi trains from, with MCAP, HDF5 and RLDS on request.
Model Inputs

What π0 and π0.5 Take as Input

Physical Intelligence's π0 paper defines each observation as multiple RGB images, a language command and the robot's proprioceptive state, and the model predicts a chunk of future actions with conditional flow matching. Pi0 fine-tuning data carries each of those streams at every step.

Camera images

2 or 3 images per robot in the paper. Its bimanual ARX and AgileX setups use two wrist cameras and one base camera.

Language command

A plain-language task instruction, read with the images and robot state.

Proprioceptive state

The robot's joint configuration, zero-padded with the actions to the largest robot's size, 18 dimensions in the paper.

Action chunks

Chunks of H=50 actions, which let π0 control robots at up to 50 Hz on dexterous tasks such as laundry folding.

π0.5 extends π0 for open-world generalization. Its paper feeds the model images from all cameras, the task prompt and the robot's configuration, with the state discretized into text tokens. It predicts a high-level subtask in text, then a flow-matching action chunk that Physical Intelligence describes as 50 steps, or one second, of joint actions.

Compare the Fields

π0 Observation and Action Fields Against Each Episode

How the fields in Physical Intelligence's papers line up with a NeuralChainAI teleoperation episode.

FieldWhat π0 and π0.5 readWhat each episode provides
Camera imagesπ0: 2 or 3 RGB images per robot, such as two wrist cameras and a base camera. π0.5: all cameras, four on its mobile manipulatorsWrist and scene views on bimanual arms, or head and wrist views on the G1, agreed per program
Language commandπ0: a language command. π0.5: the task prompt, plus a predicted high-level subtaskLanguage task labels written to your task list, added on request
Proprioceptive stateπ0: joint configuration, zero-padded to the largest robot's size. π0.5: the configuration as text tokensJoint states per arm with gripper or hand state, plus leg joints in G1 walking episodes
ActionsChunks of 50 continuous actions from flow matching, at up to 50 Hz in the π0 paperCommanded actions at every step, on one clock with the cameras and joint states
Dataset formatLeRobot format, which the openpi README trains fromLeRobot by default, with MCAP, HDF5 and RLDS on request

See our teleoperation and capture services for every embodiment we record, and human demonstrations for robot policies for first-person capture.

Training Path

How openpi Trains π0 on Your Own Episodes

openpi is Physical Intelligence's open repository for π0, π0-FAST and π0.5, with base checkpoints pre-trained on 10k+ hours of robot data, per the README. Its fine-tuning path starts from a LeRobot dataset.

Convert to LeRobot format

Step one in the README converts your data to a LeRobot dataset, which openpi uses for training.

Map the data to the model

A data config processes the episodes, and input and output transforms map your cameras and joints to the model and back.

Compute normalization statistics

openpi computes normalization statistics for the training data before the run.

Fine-tune a base checkpoint

A training config sets hyperparameters, data config and weight loader. LoRA needs more than 22.5 GB of GPU memory and full fine-tuning more than 70 GB.

The README notes that π0 was developed for Physical Intelligence's own robots, which differ from platforms such as ALOHA and DROID, so episodes from the robot you deploy carry the fine-tune. Our physical AI and robotics consulting team can help with the data config and training run. The repository is licensed Apache 2.0, with a separate Gemma license file.

Post-Training Data

How Much Post-Training Data π0 Uses

Physical Intelligence pre-trains π0 on a large, diverse corpus, then fine-tunes it on narrower, carefully curated data, with these figures for that second stage:

1 The π0 paper: only 5 hours for the simplest tasks, and 100 or more hours for the most complex.
2 The openpi announcement: between 1 and 20 hours was sufficient to fine-tune to a variety of tasks.
3 The π0.5 paper uses the same two stages, pre-training then post-training.
Sizing Your Pilot

We size the pilot to your task.

Episode volume depends on task difficulty, the variations your evaluation covers and the embodiment. We scope the pilot against those, and scaled collection follows your evaluation results.

Pilot sized per task and variation
Human-verified episodes checked against your criteria
Scale set by your evaluation results
Robots and Hands

Robots That Fit openpi Post-Training

We record episodes for π0 and π0.5 on these robots, so the fine-tune learns your embodiment.

Bimanual tabletop arms

YAM and OpenArm-class arms recorded together, with joint states and actions for both arms and a wrist view from each.

Other arms on request

ARX, PiPER and SO-101 arms. The π0 paper itself trains on bimanual ARX and AgileX setups.

Unitree G1, stationary

The G1 at a counter or table, with head and wrist views and upper-body and hand joint states.

Unitree G1, walking tasks

Walking and manipulation in one episode, with leg joint states. Our GR00T page covers the G1 for NVIDIA's model.

Grippers and hands on the G1

Dex1 grippers and tactile hands, plus Dex3, Inspire or BrainCo Revo 2 hands, set the hand dimensions in each vector.

UMI handheld

Ready-made UMI sets or new handheld collection, with the mapping to your arm agreed before the pilot.

Delivery and Licensing

LeRobot Delivery, Ready-Made Sets and Licensing

Episodes ship in LeRobot format, which openpi trains from. Ready-made sets cover UMI, the G1 and YAM across kitchen, bedroom and staged retail tasks, plus an OpenArm bimanual set. Each delivery can carry:

1 LeRobot format by default, with MCAP, HDF5 and RLDS on request.
2 Add-on tactile or force-torque streams, and depth on bimanual arms and UMI.
3 Provenance and consent records, with signed agreements permitting commercial training.
Licensing

Ready-made or custom, on the terms you choose.

Ready-made data is licensed non-exclusive and can be sublicensed to your own client. Custom collection, a new task on a robot we record with or a new robot or embodiment, is non-exclusive or exclusive. Exclusivity covers the data and its resale rights, separate from the robot. Exclusive terms are priced above non-exclusive, and term and scope are agreed once a pilot is defined.

Ready-made sets, non-exclusive and sublicensable
Custom work on a new task or a new embodiment
Term and scope agreed once a pilot is defined
Scoping

What Shapes the Engagement

We scope every π0 data program against these factors with you, then quote it as a whole.

1

Model and checkpoint

π0, π0-FAST or π0.5, and the base checkpoint you fine-tune from.

2

Embodiment and end effector

Bimanual arms, the G1 or UMI, with the gripper or hand fitted to your objects.

3

Camera views

Wrist, scene or head views matched to your deployed policy.

4

Tasks and environments

Ready-made kitchen, bedroom and staged retail families, or your own evaluation tasks.

5

Streams, labels and format

Joint states, actions, language labels and add-on streams, in LeRobot or an agreed format.

6

Acceptance and QC criteria

Rejects such as a dropped object, a collision, a stream out of sync or an incomplete task.

7

Pilot size

Episodes and task variations in the pilot, enough to fine-tune and evaluate on.

8

Exclusivity and sublicense rights

Non-exclusive, sublicensable ready-made data, or custom collection on either term.

Start Today

Have a π0 Task to Record?

Send us the task, the robot and the camera views your policy reads. We'll scope the pilot and send a representative sample pack with the same modalities, calibration and delivery format you'd get in production, so your team can inspect sensor alignment before any commitment.

Book a Strategy Session →
Ask us about
Episodes for π0, π0-FAST or π0.5 post-training
YAM and OpenArm-class arms, with ARX, PiPER and SO-101 on request
Standing and walking episodes on the Unitree G1
Ready-made UMI, G1 and YAM sets and the OpenArm set
LeRobot delivery, with MCAP, HDF5 and RLDS on request
Questions

Frequently Asked Questions on Pi0 Fine-Tuning Data

The π0 paper reports only 5 hours for the simplest tasks and 100 or more hours for the most complex. We size the pilot against your task and evaluation, then scale from what it shows.

LeRobot by default, which the openpi README trains from, with MCAP, HDF5 and RLDS on request. openpi's data config and transforms map your cameras and joints to the model, and our team can help set them up.

Yes, on request. π0 reads a language command with its images and robot state, per the paper, so we write language task labels to your task list and add them to each episode.

Ready-made data is non-exclusive and can be sublicensed to your own client. Custom collection can be exclusive, covering the data and its resale rights, and exclusive terms are priced above non-exclusive.

Scope Your Pi0 Fine-Tuning Data

A 30-minute strategy call. Tell us the robot, camera views and tasks. We'll scope a pilot for π0 or π0.5 post-training with the streams, format and acceptance criteria your team needs.

Discuss Your Project