Physical IntelligenceTranslation site

Company Overview

Physical Intelligence is a U.S.-based robotics and AI company founded in 2024 and headquartered in San Francisco, with Karol Hausman serving as CEO. The team brings together researchers in robotics and machine learning with the goal of building a reusable layer of robotic intelligence that enables different hardware to perform a variety of tasks using the same type of model. The company was named to Forbes’ 2026 AI 50 list.

Its focus is on models, learning algorithms, and real-world data, rather than on selling a specific type of humanoid robot as its sole product. For robot manufacturers and deployment teams, this type of technology is expected to reduce the burden of having to redesign control software for each task.

Milestones

The team has successively published research on models such as π0, π0.5, and π0.6, continuing to explore how to combine knowledge learned from the internet, robot demonstrations, and practical execution experience. A collaboration case study released in February 2026 showcased the models’ applications in folding laundry and packing e-commerce orders, with partners including Weave and Ultra.

The π0.7 study, published in 2026, further emphasizes the use of language, sub-goal images, and task metadata to guide robots. The paper demonstrates task generalization across environments and robot configurations; however, these findings do not imply that robots are already capable of unconditionally performing any real-world task.

Products & Services

• Basic Robot Models:The π series models integrate visual perception, text commands, and motor output, providing a foundation of general-purpose capabilities for tasks such as grasping, sorting, folding, and packaging.
• Collaboration on Model Adaptation and Deployment:Collaborate with robotics companies to collect and utilize scenario data, adjust models based on actual equipment, workstations, and operational processes, and evaluate the effectiveness of the implementation.
• Published Research Findings:This technical paper introduces model architecture, training methods, and experiments to help researchers understand multi-task robot learning and cross-hardware transfer.
• Real-World Task Verification:Test the model using the partner’s existing robots and in real-world customer scenarios, evaluating factors such as task completion, human intervention, and operational reliability.

core technology

Its core approach is the Vision-Language-Action (VLA) model. The model must both understand goals such as “folding clothes” and process changes in position, shape, and contact within the camera’s field of view, ultimately generating continuous actions. Early π0 research combined pre-trained vision-language capabilities with flow-matching action generation, bridging the gap between high-level semantics and fine-grained control.

π0.7 further leverages rich contextual information by incorporating not only the task text but also images of sub-goals and task performance data, while combining both robotic and non-robotic data. The significance of this approach lies in enabling the model to learn “how to perform” a task, rather than merely recognizing “what to do.” Real-world deployment still requires coordination with hardware adaptation, fault handling, and manual takeover mechanisms.

Take folding clothes as an example: robots must not only recognize garments but also handle the deformation of soft materials, grasping failures, and variations in size; order packaging, meanwhile, involves dealing with product differences and coordinating with workstation equipment. Official case studies of collaborations have retained mechanisms for human intervention, indicating that current applications still require integrating model capabilities with complete operational workflows. When evaluating such systems, one should consider throughput, failure recovery, and the workload on staff collectively—rather than focusing solely on whether individual actions appear smooth.

development prospect

Based on existing research and collaboration cases, Physical Intelligence is poised to become a source of reusable models for robotics companies, enabling them to gradually roll out capability upgrades across different devices. Scenarios with relatively well-defined task boundaries—such as warehouse packaging and commercial laundry—may be better suited for evaluating actual benefits.

This assessment remains subject to factors such as data collection costs, rare failures, operational safety, and deployment and maintenance conditions. The physical environment in which robots operate is constantly changing, and experimental videos or one-time success rates are insufficient to replace long-term operational validation. What truly matters is whether the robots can operate reliably at different customer sites and whether model updates can reduce the need for human intervention without introducing new risks.