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 in the PaLI configuration with prefixed visual tokens from a ViT to Gemma 2B
$ pip install pi-zero-pytorchimport 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)@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{Yao2024FasterDiTTF,
title = {FasterDiT: Towards Faster Diffusion Transformers Training without Architecture Modification},
author = {Jingfeng Yao and Wang Cheng and Wenyu Liu and Xinggang Wang},
year = {2024},
url = {https://api.semanticscholar.org/CorpusID:273346237}
}