pi-zero-pytorch 0.1.24

π0 in Pytorch

pi-zero-pytorch (wip)

Implementation of π₀ the robotic foundation model architecture proposed by Physical Intelligence

Summary of this work would be that it is a simplified Transfusion (Zhou et al.) with influence from Stable Diffusion 3 (Esser et al.), mainly the adoption of flow matching instead of diffusion for policy generation, as well as the separation of parameters (Joint Attention from mmDIT). They build on top of a pretrained vision language model, PaliGemma 2B.

Update: The official repository has been open sourced!

Appreciation

• Einops for the amazing pack and unpack, used extensively here for managing various token sets

• Flex Attention for allowing for easy mixture of autoregressive and bidirectional attention

• @Wonder1905 for the code review and identifying issues

• You? maybe a phd student who wants to contribute to the latest SOTA architecture for behavioral cloning?

Install

$ pip install pi-zero-pytorch

Usage

import torch
from pi_zero_pytorch import π0

model = π0(
    dim = 512,
    dim_action_input = 6,
    dim_joint_state = 12,
    num_tokens = 20_000
)

vision = torch.randn(1, 1024, 512)
commands = torch.randint(0, 20_000, (1, 1024))
joint_state = torch.randn(1, 12)
actions = torch.randn(1, 32, 6)

loss, _ = model(vision, commands, joint_state, actions)
loss.backward()

# after much training

sampled_actions = model(vision, commands, joint_state, trajectory_length = 32) # (1, 32, 6)

To do online learning, you need to make sure policy_optimizable is set to True when instantiating the model. Then do the following

from pi_zero_pytorch import π0, Agent, EPO

model = π0(
    ...
    policy_optimizable = True
)

# you'll want to supply your own environment

from pi_zero_pytorch.mock_env import Env
mock_env = Env((256, 256), 2, 32, 1024, 12)

# wrap the model with `Agent`, which will instantiate actor critic for PPO

agent = Agent(model)

# pass your agent and environment to EPO for learning to be orchestrated

epo = EPO(agent, mock_env)

# gather memories from environment

memories = epo.gather_experience_from_env(steps = 10)

# learn from memories

epo.learn_agent(memories, batch_size = 2)

Contributing

At the project root, run

$ pip install '.[test]' # or `uv pip install '.[test]'`

Then add your tests to tests/test_pi_zero.py and run

$ pytest tests/

That's it

Citation

@misc{Black2024,
    author  = {Kevin Black, Noah Brown, Danny Driess, Adnan Esmail, Michael Equi, Chelsea Finn, Niccolo Fusai, Lachy Groom, Karol Hausman, Brian Ichter, Szymon Jakubczak, Tim Jones, Liyiming Ke, Sergey Levine, Adrian Li-Bell, Mohith Mothukuri, Suraj Nair, Karl Pertsch, Lucy Xiaoyang Shi, James Tanner, Quan Vuong, Anna Walling, Haohuan Wang, Ury Zhilinsky},
    url     = {https://www.physicalintelligence.company/download/pi0.pdf}
}

@inproceedings{Zhou2024ValueRL,
    title   = {Value Residual Learning For Alleviating Attention Concentration In Transformers},
    author  = {Zhanchao Zhou and Tianyi Wu and Zhiyun Jiang and Zhenzhong Lan},
    year    = {2024},
    url     = {https://api.semanticscholar.org/CorpusID:273532030}
}

@inproceedings{Darcet2023VisionTN,
    title   = {Vision Transformers Need Registers},
    author  = {Timoth'ee Darcet and Maxime Oquab and Julien Mairal and Piotr Bojanowski},
    year    = {2023},
    url     = {https://api.semanticscholar.org/CorpusID:263134283}
}

@article{Li2024ImmiscibleDA,
    title   = {Immiscible Diffusion: Accelerating Diffusion Training with Noise Assignment},
    author  = {Yiheng Li and Heyang Jiang and Akio Kodaira and Masayoshi Tomizuka and Kurt Keutzer and Chenfeng Xu},
    journal = {ArXiv},
    year    = {2024},
    volume  = {abs/2406.12303},
    url     = {https://api.semanticscholar.org/CorpusID:270562607}
}

@inproceedings{Sadat2024EliminatingOA,
    title   = {Eliminating Oversaturation and Artifacts of High Guidance Scales in Diffusion Models},
    author  = {Seyedmorteza Sadat and Otmar Hilliges and Romann M. Weber},
    year    = {2024},
    url     = {https://api.semanticscholar.org/CorpusID:273098845}
}

@article{Bulatov2022RecurrentMT,
    title   = {Recurrent Memory Transformer},
    author  = {Aydar Bulatov and Yuri Kuratov and Mikhail S. Burtsev},
    journal = {ArXiv},
    year    = {2022},
    volume  = {abs/2207.06881},
    url     = {https://api.semanticscholar.org/CorpusID:250526424}
}

@inproceedings{Bessonov2023RecurrentAT,
    title   = {Recurrent Action Transformer with Memory},
    author  = {A. B. Bessonov and Alexey Staroverov and Huzhenyu Zhang and Alexey K. Kovalev and D. Yudin and Aleksandr I. Panov},
    year    = {2023},
    url     = {https://api.semanticscholar.org/CorpusID:259188030}
}

@article{Zhu2024HyperConnections,
    title   = {Hyper-Connections},
    author  = {Defa Zhu and Hongzhi Huang and Zihao Huang and Yutao Zeng and Yunyao Mao and Banggu Wu and Qiyang Min and Xun Zhou},
    journal = {ArXiv},
    year    = {2024},
    volume  = {abs/2409.19606},
    url     = {https://api.semanticscholar.org/CorpusID:272987528}
}

@inproceedings{Sun2025F5RTTSIF,
    title   = {F5R-TTS: Improving Flow-Matching based Text-to-Speech with Group Relative Policy Optimization},
    author  = {Xiaohui Sun and Ruitong Xiao and Jianye Mo and Bowen Wu and Qun Yu and Baoxun Wang},
    year    = {2025},
    url     = {https://api.semanticscholar.org/CorpusID:277510064}
}

@inproceedings{Wang2025EvolutionaryPO,
    title = {Evolutionary Policy Optimization},
    author = {Jianren Wang and Yifan Su and Abhinav Gupta and Deepak Pathak},
    year  = {2025},
    url   = {https://api.semanticscholar.org/CorpusID:277313729}
}

dear alice