π₀.₅ (Pi05) (LeRobot)

π₀.₅ (Pi05) (LeRobot)

A robot control model that reads camera images, robot state and an optional instruction and outputs continuous actions. It is a base checkpoint for researchers who fine-tune on their own robot data.

Made by lerobot on Hugging Face. These notes are Mindket's, written from its model page and model card; the model's own files and license are on Hugging Face.

About

What it is: Pi05 is the LeRobot version of π₀.₅, a vision-language-action model from Physical Intelligence. It reads camera images, the robot's state and an optional language instruction, then outputs continuous robot actions. This checkpoint has 3.6 billion parameters, is built on Google's PaliGemma, and is meant as a base you fine-tune on your own task. Only the flow-matching action head is included. The original work's subtask prediction, action tokenization and RL parts are not.

Who it suits: It suits robotics researchers and machine learning engineers who have demonstration data and want to fine-tune a policy. People with a hobby robot arm can use it to run a trained policy and record evaluation episodes.

Title: Who might use it
Industry | Who uses it | How they use it | Example
Robotics research | University and lab teams | Fine-tune the base checkpoint on their own demonstrations | Train a policy on a dataset of recorded episodes
Hobby robotics | Builders with a low-cost arm | Run a trained policy and record evaluation runs | Record 10 evaluation episodes with the record script
Machine learning | ML engineers | Test inference on a frame from a dataset | Load a frame from the LIBERO dataset and call select_action

If you want a model that completes a task with no training, you can skip it, because the card calls it a base model.

Where to use it: The model runs through the LeRobot library in Python. The card's recording example drives a physical arm with cameras attached.

• Python with the LeRobot library, on Windows, macOS or Linux
• A robot with cameras, through the lerobot-record script
• A GPU for training, since the card's training command sets the device to cuda

Title: Jobs it does
Use case | Who | What they get back
Predict an action from a dataset frame | Developers | A predicted action for that frame
Fine-tune on your own data | Researchers | A policy checkpoint trained on your dataset
Evaluate on a robot | Robot builders | Recorded evaluation episodes

How to set it up: You need Python, the LeRobot library with its pi extra, and PyTorch. Video datasets may also need ffmpeg for torchcodec.

• Python with pip
• The LeRobot library with the pi extra
• PyTorch, which the sample code imports
• Optional video tools such as ffmpeg for torchcodec

1. Install LeRobot with the pi extra. The card's own line pulls it from the project's GitHub main branch.
2. Optionally fetch the checkpoint files with the Hugging Face tools.
3. Run the sample below on one dataset frame, which needs no robot.

pip install -U huggingface_hub
hf download lerobot/pi05_base

How to use it: Load the policy, take one frame from a dataset, preprocess it, and call select_action. The postprocessor then converts the action for use. To train, call lerobot-train with your dataset and the base checkpoint.

import torch
from lerobot.datasets.lerobot_dataset import LeRobotDataset
from lerobot.policies.factory import make_pre_post_processors

# Swap this import per-policy
from lerobot.policies.pi05 import PI05Policy

# load a policy
model_id = "lerobot/pi05_base"  # <- swap checkpoint
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")

policy = PI05Policy.from_pretrained(model_id).to(device).eval()

preprocess, postprocess = make_pre_post_processors(
    policy.config,
    model_id,
    preprocessor_overrides={"device_processor": {"device": str(device)}},
)
# load a lerobotdataset (we will replace with a simpler dataset)
dataset = LeRobotDataset("lerobot/libero")

# pick an episode
episode_index = 0

# each episode corresponds to a contiguous range of frame indices
from_idx = dataset.meta.episodes["dataset_from_index"][episode_index]
to_idx   = dataset.meta.episodes["dataset_to_index"][episode_index]

# get a single frame from that episode (e.g. the first frame)
frame_index = from_idx
frame = dict(dataset[frame_index])

batch = preprocess(frame)
with torch.inference_mode():
    pred_action = policy.select_action(batch)
    # use your policy postprocess, this post process the action
    # for instance unnormalize the actions, detokenize it etc..
    pred_action = postprocess(pred_action)

lerobot-train \
  --dataset.repo_id=${HF_USER}/<dataset> \
  --output_dir=./outputs/[RUN_NAME] \
  --job_name=[RUN_NAME] \
  --policy.repo_id=${HF_USER}/<desired_policy_repo_id> \
  --policy.path=lerobot/[BASE_CHECKPOINT] \
  --policy.dtype=bfloat16 \
  --policy.device=cuda \
  --steps=100000 \
  --batch_size=4

Title: Main tools
Tool | What it does
PI05Policy.from_pretrained | Loads the checkpoint onto your device
make_pre_post_processors | Builds the steps that prepare inputs and convert outputs
policy.select_action | Predicts an action from a prepared batch
lerobot-train | Trains or fine-tunes a policy on a dataset
lerobot-record | Runs a policy on a robot and records evaluation episodes

Example: You load the checkpoint, pick the first frame of episode 0 from the LIBERO dataset, and run it through the preprocessor. Calling select_action returns a predicted action, which the postprocessor converts into values a robot controller can use.

For newcomers and experts: If you are new, a vision-language-action model looks at camera pictures and an optional instruction, then decides what a robot should do next. The first small thing to try is the sample above on one dataset frame. It needs no robot.

If you know the field, the model takes multi-view images and proprioceptive state and predicts continuous actions, trained with a flow-matching objective. It follows the original reference code for compatibility. The trade-off is scope: it is a base checkpoint with only part of the original work, so expect to fine-tune.

Camera images, state and an instruction come in
The preprocessor prepares the batch
The policy predicts a continuous action
The postprocessor converts the action

Good to know: The checkpoint is large. The 4-bit figure below is an estimate from the Hugging Face record, not a file you download.

Title: Approximate weight size
F32: 14.5 GB
4-bit estimate: 1.8 GB

• License: Gemma
• Cost to run: none named, though the training example sets a GPU and 100000 steps
• Data it reaches: datasets and files from Hugging Face, and with the record script a robot and its cameras
• Not for: use without fine-tuning, or for the subtask prediction, action tokenization and RL parts of the original work

On Hugging Face, 13 November 2026

• Task: robotics
• Library: lerobot
• Parameters: 3.6B
• Weights: about 14.5 GB in F32; about 1.8 GB at 4-bit (estimate)
• Formats: Safetensors
• Downloads in the last 30 days: 21,996
• Likes: 119
• License: gemma
• Languages: en
• Based on: google/paligemma-3b-pt-224
• Quantized versions: 1
• Fine-tunes: 860
• Tags: lerobot, safetensors, vision-language-action, imitation-learning, robotics

Links

• Model: https://huggingface.co/lerobot/pi05_base
• Files: https://huggingface.co/lerobot/pi05_base/tree/main

Get the model

Download its files with Hugging Face's command line:

pip install -U huggingface_hub
hf download lerobot/pi05_base

Then follow the usage in its model card. Nothing in this kit runs by itself.