Blogs›Physical Intelligence›Steerable Generalist Robotic Models
The release of the openpi repository provides code and model weights for the π0 general-purpose robotic foundation model, enabling fine-tuning for various robots and tasks. It includes checkpoints fine-tuned for specific platforms like ALOHA and DROID, example code for inference and fine-tuning, and a PyTorch port. This release aims to foster community experimentation and development in embodied AI.
4/16/2026
Introduced π0.7, a general-purpose robotic model exhibiting compositional generalization. Key technical contributions include: training with diverse multimodal prompts (language, metadata, control modalities, visual subgoals) to specify task execution 'how-to', enabling the model to leverage broad data sources and recombine skills for novel tasks. Demonstrated specialist-level dexterity and cross-embodiment transfer without fine-tuning.
12/22/2025
This post details the application of the π0.7 model to a new set of physical manipulation tasks, referred to as the "Robot Olympics." It demonstrates the model's ability to generalize and learn new skills with fine-tuning, achieving success rates on tasks like opening a self-closing door, turning a sock inside-out, using a key, peeling an orange, and cleaning a greasy pan. The post contrasts the performance of the fine-tuned π0.7 model with a baseline VLM, emphasizing the necessity of large-scale robot pre-training. It also discusses the challenges of physical intelligence and how multimodal models with diverse real-world data can overcome Moravec's paradox.
12/16/2025
Demonstrates emergent human-to-robot transfer in VLAs by fine-tuning with human video data, achieving a 2x performance improvement. Investigates the correlation between pre-training data scale/diversity and the effectiveness of human data transfer, showing that increased robot pre-training data improves the model's ability to absorb human data. Visualizes latent embeddings to show emergent alignment between human and robot data representations as pre-training scales.
11/17/2025
Introduces Recap (RL with Experience & Corrections via Advantage-conditioned Policies), a method for training robotic models that combines demonstrations, expert corrections, and autonomous experience. Details the use of a value function to predict task progress and an advantage-conditioned policy for learning from experience. Demonstrates improvements in throughput and failure rates for VLA models performing tasks like espresso making, box assembly, and laundry folding.
2/4/2025
This post announces the open-sourcing of the π0 robotic foundation model via the openpi repository. It provides access to the base pre-trained model, fine-tuned checkpoints for platforms like ALOHA and DROID, example inference and fine-tuning code, and a PyTorch port. The release aims to enable broader experimentation and adaptation of the model to diverse robotic systems and tasks.