Skip to content
ContentLora

    Tip: press / anywhere to search.

    organization

    Physical Intelligence

    Also known as π, Pi, Physical Intelligence Inc.

    Physical Intelligence is a startup that builds general-purpose "robot brains": vision-language-action models meant to run on many kinds of robots.[1][2] Its π0.5 model cleaned kitchens and bedrooms in homes it had never seen.[3] In March 2026 it was reported to be raising about $1 billion at a valuation above $11 billion.[4]

    Editor reviewedUpdated Robotics and embodied AIArtificial intelligence
    Key facts

    Physical Intelligence is a startup trying to build a single AI model that can control many different robots, and it has attracted multibillion-dollar valuations for the effort. Its π (“pi”) models are vision-language-action models: they take images and language instructions and output robot actions.[1][4] It was founded by former DeepMind staff and others.[5] Most performance figures below come from the company’s own papers and blog posts.

    Models

    π0 (October 2024) puts a flow-matching “action expert” on top of a pretrained vision-language model. Flow matching is a generative technique related to diffusion.[1] It was trained on data from single-arm, dual-arm and mobile robots, and evaluated on tasks such as folding laundry, cleaning tables and assembling boxes.[2]

    FAST (January 2025) is a way of turning continuous robot motions into tokens a language-model-style network can predict. It compresses action sequences with the discrete cosine transform. Combined with π0, the company reports it matched diffusion-based models on 10,000 hours of robot data while cutting training time by up to five times.[6]

    π0.5 (April 2025) aimed at open-world generalization. It was co-trained on varied data, including object detections, high-level subtask predictions and low-level actions. The paper shows it carrying out long, multi-step tasks such as cleaning a kitchen or bedroom in homes it had not been trained in.[3]

    π*0.6 (November 2025) added practice. A reinforcement-learning method called Recap trains the model on demonstrations, then on human corrections, then on its own autonomous attempts. The company says this more than doubled throughput on some of the hardest tasks.[7] Its demonstrations included making espresso drinks from 5:30am to 11:30pm and assembling 59 chocolate-packaging boxes in a real factory.[8]

    Memory (March 2026). A system called MEM lets the models follow tasks up to fifteen minutes long, keeping short-term memories as raw observations and long-term ones as short text notes.[9]

    π0.7 (April 2026) is the latest model. The company says it performs dexterous tasks as well as its fine-tuned specialists and shows first signs of compositional generalization, such as folding laundry on a robot that had no laundry-folding data.[10] It credits data from many robots, human video and autonomous runs, plus prompts that describe how a task should be done, not only what to do.[11]

    Open source

    In February 2025 the company released the code and weights for π0 in its openpi repository.[12] The repository now also holds π0-FAST and π0.5, with base checkpoints pre-trained on more than 10,000 hours of robot data.[12] In its own experiments, 1 to 20 hours of data was enough to fine-tune π0 to a variety of tasks.[13]

    Approach

    The company’s research centres on cross-embodiment training. One model learns from many robot types instead of one model per robot.[2] The academic Open X-Embodiment project tested the same idea and found that pooled multi-robot data produces positive transfer.[14] Physical Intelligence also works with companies that deploy robots. One partner, Weave, ran π0.6 folding laundry in a San Francisco laundromat.[15]

    Funding

    Before 2026 the company was valued at $5.6 billion. In March 2026 Bloomberg News reported it was in talks to raise about $1 billion at a valuation above $11 billion.[4] As of October 2026 this page has not confirmed whether that round closed.

    Competition

    Its main rivals in general-purpose robot models are big-tech efforts such as Google DeepMind‘s Gemini Robotics[16] and NVIDIA‘s open Isaac GR00T models.[17] Humanoid makers such as Figure build their own models for their own robots.[18]

    Questions readers ask

    What is π0?

    π0 is Physical Intelligence's first generalist robot model. It adds a flow-matching action generator on top of a pretrained vision-language model.[1]

    Does Physical Intelligence build robots?

    Its published work centres on models trained on data from many robot types, including single-arm, dual-arm and mobile manipulators.[2]

    Are Physical Intelligence's models open source?

    Partly. Its openpi repository provides code and weights for π0, π0-FAST and π0.5, with base checkpoints pre-trained on more than 10,000 hours of robot data. Its newer models, π*0.6 and π0.7, are described in papers and blog posts.[12][7][10]

    Are its models used outside the lab?

    The company says it works with robot deployers such as Weave, whose robots ran π0.6 folding laundry at a San Francisco laundromat.[15]

    How much is Physical Intelligence worth?

    Its previous valuation was $5.6 billion. Bloomberg reported in March 2026 that it was in talks to raise about $1 billion at more than $11 billion; as of October 2026 this page has not confirmed that the round closed.[4]

    Sources

    Each numbered claim is a statement we checked against the sources listed with it. Status shows how well established it is.

    1. [1]

      Physical Intelligence's π0 (October 2024) is a vision-language-action model that adds a flow-matching action generator on top of a pretrained vision-language model. confirmedas of 2024-10-31

    2. [2]

      π0 was trained on data from multiple robot types, including single-arm, dual-arm and mobile manipulators, and evaluated on tasks such as laundry folding, table cleaning and box assembly. confirmedas of 2024-10-31

    3. [3]

      Physical Intelligence's π0.5 (April 2025) uses co-training on heterogeneous tasks and multimodal examples (images, language commands, object detections, semantic subtask predictions and low-level actions) to perform long-horizon tasks such as cleaning a kitchen or bedroom in homes not seen during training. confirmedas of 2025-04-22

    4. [4]

      In March 2026 Bloomberg News reported that Physical Intelligence was in talks to raise about $1 billion at a valuation above $11 billion, roughly double its previous $5.6 billion valuation. reportedas of 2026-03-29

    5. [5]

      Physical Intelligence was founded by former DeepMind staff and develops AI software that lets robots learn tasks such as folding clothes and assembling boxes. reportedas of 2026-03-29

    6. [6]

      Physical Intelligence's FAST tokenizer (January 2025) compresses robot actions with the discrete cosine transform; combined with π0 it matched diffusion VLAs on 10,000 hours of robot data while cutting training time by up to five times. confirmedas of 2025-01-16

    7. [7]

      In November 2025 Physical Intelligence introduced π*0.6, trained with a reinforcement-learning method called Recap that combines demonstrations, corrections and autonomous experience; the company says it more than doubled throughput on some of the hardest tasks. confirmedas of 2025-11-17

    8. [8]

      Physical Intelligence reported that π*0.6 made espresso drinks from 5:30am to 11:30pm, folded 50 novel laundry items in a new home and assembled 59 chocolate-packaging boxes in a real factory. confirmedas of 2025-11-17

    9. [9]

      Physical Intelligence's Multi-scale Embodied Memory (MEM), published in March 2026, lets its models track tasks up to fifteen minutes long, keeping short-term memories as observations and long-term memories as natural-language summaries. confirmedas of 2026-03-03

    10. [10]

      Physical Intelligence said in April 2026 that its π0.7 model performs dexterous tasks as well as fine-tuned specialists and shows first signs of compositional generalization, such as folding laundry on a robot with no laundry-folding data. confirmedas of 2026-04-16

    11. [11]

      Physical Intelligence attributes π0.7's generalization to diverse data from many robots, human data and autonomous episodes, combined with richer multimodal prompts that specify how a task should be done. confirmedas of 2026-04-16

    12. [12]

      In February 2025 Physical Intelligence released the code and weights for π0, and its openpi repository later added the autoregressive π0-FAST and the π0.5 model, with base checkpoints pre-trained on more than 10,000 hours of robot data. confirmedas of 2026-10-10

    13. [13]

      Physical Intelligence said 1 to 20 hours of data was enough to fine-tune π0 to a variety of tasks in its own experiments. confirmedas of 2025-02-04

    14. [14]

      The RT-X models trained on Open X-Embodiment (paper first posted October 2023) showed positive transfer, improving the capabilities of multiple robots by using experience from other robot platforms. confirmedas of 2023-10-13

    15. [15]

      In February 2026 Physical Intelligence said it works with companies that deploy robots, including Weave, whose robots ran π0.6 folding laundry at a San Francisco laundromat, and Ultra. confirmedas of 2026-02-24

    16. [16]

      On July 30, 2026 Google DeepMind announced Gemini Robotics 2, a vision-language-action model it says can control full humanoids, including walking and manipulation, as well as bi-arm robots; previous models in the family controlled a humanoid's upper body for tabletop tasks. confirmedas of 2026-07-30

    17. [17]

      On March 16, 2026 NVIDIA made GR00T N1.7 available in early access with commercial licensing and previewed GR00T N2, which it says succeeds at new tasks in new environments more than twice as often as leading VLA models and is slated for release by the end of 2026. confirmedas of 2026-03-16

    18. [18]

      Figure says Helix 02 (January 2026) controls the whole robot from pixels, unloading and reloading a dishwasher across a kitchen in a four-minute autonomous task, using a learned whole-body controller, System 0, trained on more than 1,000 hours of human motion data and sim-to-real reinforcement learning. confirmedas of 2026-01-27

    Revision history (2)
    1. Page created.
    2. Expanded: open-sourcing (openpi), the FAST tokenizer, π*0.6 and Recap, memory (MEM), π0.7 and deployment partners, all from the company's own papers and posts.

    Created Oct 10, 2026. Last reviewed by an editor on Oct 10, 2026. Next scheduled review: Jan 10, 2027.

    Cite this page

    "Physical Intelligence." ContentLora, updated Oct 10, 2026. https://contentlora.com/wiki/physical-intelligence

    Spotted an error? Suggest a correction or emailcorrections@contentlora.com.

    Keep exploring

    Connections
    Open X-Embodiment Related cross-embodiment …Gemini Robotics Competes withNVIDIA Isaac GR00T Competes withGoogle DeepMind Founders include former s…
    ☰ Contents