Steerable Generalist Robotic Models

5 posts · Feb 2025 – Apr 2026

What this trail covers

The π0 foundation model continues to evolve, demonstrating enhanced generalist capabilities in physical manipulation tasks. The latest iteration, π*0.6, is trained using Recap (RL with Experience & Corrections via Advantage-conditioned Policies), which combines demonstrations, expert corrections, and autonomous experience. This approach significantly improves robustness and throughput on complex real-world tasks like making espresso drinks, assembling boxes, and folding laundry, enabling practical levels of reliability and efficiency.

All Physical Intelligence trails

Apr 2026A Steerable Model with Emergent CapabilitiesIntroduced π0.7, a steerable general-purpose robotic foundation model that demonstrates compositional generalization. This model can perform a wide range of dexterous tasks with specialist-level performance without fine-tuning, and can solve novel tasks by recombining learned skills. Generalization is achieved through training with diverse multimodal prompts (language, metadata, control modalities, visual subgoals) that specify task execution details, enabling the integration of broader and more diverse datasets.
Dec 2025Moravec's Paradox and the Robot OlympicsThis post details the fine-tuning of the π0.6 robotic foundation model to address a set of 'Robot Olympics' manipulation tasks. It demonstrates initial solutions for 3 out of 5 proposed categories, achieving 'gold medal' status in tasks like opening a self-closing door, using a key, and cleaning a greasy pan. The contribution lies in showcasing the model's ability to generalize to novel, everyday physical tasks with under 9 hours of data collection per task, significantly outperforming a baseline VLM. The post also discusses the challenges of physical intelligence and the importance of large-scale robotic pre-training for achieving such capabilities.
Dec 2025Emergence of Human to Robot Transfer in Vision-Language-Action ModelsThis post details the emergence of human-to-robot transfer in vision-language-action (VLA) models as a function of scaling robot training data. The authors demonstrate that simply co-finetuning a scaled-up VLA foundation model (π0.5) with human egocentric video data, without explicit transfer learning mechanisms, results in a ~2x performance improvement on generalization tasks present only in the human data. They show that this emergent transfer capability is directly correlated with the diversity and scale of the robot pre-training data, as evidenced by improved latent space alignment between human and robot representations. The contribution is the discovery and empirical validation of this emergent capability, suggesting that scaling foundation models unlocks new ways to leverage diverse data sources.
Nov 2025A VLA that Learns from ExperienceIntroduces Recap, a novel reinforcement learning method that enhances vision-language-action (VLA) models by incorporating expert corrections and autonomous experience alongside initial demonstrations. Recap trains an advantage-conditioned policy by learning a value function to assess task progress and guide learning. This method is applied to the π0.6 VLA model, resulting in π*0.6, which demonstrates more than double the throughput and reduced failure rates on complex robotic tasks, achieving practical real-world performance.
Feb 2025Open Sourcing π0This 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.
Advertisement