1. News
  2. Pi0: The Open Source Model That Improves Robotics

Pi0: The Open Source Model That Improves Robotics | Microsoft generative ai course | Generative ai in investment management | Generative ai in banking mckinsey | Turtles AI

Pi0: The Open Source Model That Improves Robotics
Physical Intelligence opens its foundation model for robotic control to the public, promising greater adaptability and innovation in the sector
Editorial Team9 February 2025

 

Physical Intelligence has open-sourced its Pi0 robotic foundation model, allowing developers to adapt it to different tasks. The openpi project includes code, weights, and checkpoints optimized for platforms such as ALOHA and DROID. The company aims to develop generalist robotic control models and collaborate with the community to improve their application.

Key Points:

  • Open source model: Physical Intelligence has released the Pi0 model code and weights, accessible via the openpi repository on GitHub.
  • Adaptability and performance: Pi0 can be fine-tuned with 1-20 hours of data for numerous tasks on robots with different configurations.
  • Optimized checkpoints: Platform-specific models such as ALOHA and DROID are available, with different control and inference strategies.
  • Collaboration and future: The openpi project aims to foster generalist robotics research and development, inspired by the evolution of open source language models.

Physical Intelligence, a San Francisco startup that has raised more than $400 million in funding, has announced the open source release of its Pi0 robotic foundation model. The technology represents a significant advance in adaptable robotics, allowing developers and researchers to customize the model’s capabilities for a wide range of practical applications. Pi0 is designed to control robots in tasks such as folding laundry, cleaning surfaces, and manipulating small objects, with the ability to be honed for specific tasks through rapid, targeted learning.

At the heart of the project is openpi, a GitHub repository that provides not only the model’s source code, but also pre-trained weights and numerous checkpoints for optimizing Pi0 on real and simulated robotic platforms. Resources included include inference scripts and tools for fine-tuning the model on existing robots. Physical Intelligence confirmed that optimizing Pi0 requires 60 to 1,200 minutes of operating data to achieve satisfactory results, which is surprisingly short compared to traditional robotic learning techniques.

The base Pi0 model was trained on a large dataset that includes several robotic platforms, including OXE and seven other systems developed by the company. Among the available variants, Pi0-FAST uses an advanced tokenizer for a more efficient autoregressive discretization, improving the model’s ability to follow linguistic instructions with greater accuracy, albeit at the cost of increased computational complexity. In addition, there are models optimized for specific environments: Pi0-FAST DROID and Pi0 DROID are specialized in controlling Franka robotic arms in diverse scenarios, while Pi0 ALOHA was designed for advanced manipulations on dual-arm robots, with tasks such as food handling or fabric folding. Finally, the Pi0 Libero checkpoint is calibrated for the Libero benchmark, providing out-of-the-box performance for this platform.

The Physical Intelligence initiative is part of a growing interest in foundation models in robotics. These models, similar to Large Language Models in the field of language, allow robots to learn new tasks with a reduced number of examples, improving generalization and reducing the need for specific data for each new task. The company emphasizes that the future of robotics will depend on the development of models capable of adapting to any machine and operating context. However, it recognizes that there are still many unsolved challenges, both in terms of scalability and integration with existing platforms.

The release of openpi therefore represents an ambitious experiment: although Pi0 was developed primarily for proprietary Physical Intelligence platforms, the company encourages the community to test it and adapt it to other configurations. Openpi is compatible with both JAX and PyTorch, thanks to a port developed in collaboration with HuggingFace, thus broadening the accessibility of the model to the research and development community. The company expects that the main use of openpi will be the refinement of Pi0 on specific robots, with the possibility of creating new implementations that further improve the adaptation and generalization capabilities.

The Robotics Summit & Expo, taking place in Boston on April 30 and May 1, will be a major opportunity to delve deeper into the evolution of robotic foundation models. Daniela Rus, director of MIT’s CSAIL, will deliver a keynote addressing physical intelligence, a concept that combines AI’s computational power with real-world interaction. The summit will feature more than 70 speakers and thousands of industry experts, providing a broad overview of current innovations and future prospects for generalist robotics.

The open source release of Pi0 marks a significant step toward more versatile and accessible physical AI, paving the way for new applications in robotic control and advanced automation.

Video