Physical Intelligence

Physical Intelligence is a robotics foundation-model company founded in 2024 and headquartered in San Francisco. It describes its goal as building general-purpose control models for physical robots, a positioning it summarizes as "the GPT of physical intelligence." Co-founder and chief executive Karol Hausman has described the company's aim as building "not just a brain for any particular robot" but "a single generalist brain that can control any robot" (Source: https://www.therobotreport.com/physical-intelligence-raises-400m-for-foundation-models-for-robotics/).

Snapshot

Valuation / Funding

Cumulative funding reported at about $1.1 billion as of the November 2025 round (Source: https://www.therobotreport.com/physical-intelligence-raises-600m-advance-robot-foundation-models/).

Overview

Physical Intelligence operates a website at pi.website and positions itself as a builder of general-purpose foundation models for robot control that are applicable across embodiments. Its stated approach is model-first and embodiment-flexible, in contrast to competitors that focus on a specific humanoid platform, such as Figure, Clone, and Tesla Optimus. A TechCrunch profile, "A peek inside Physical Intelligence, the startup building Silicon Valley's buzziest robot brains" (January 30, 2026), described the company as the most-hyped robotics AI startup in Silicon Valley and characterized its approach as treating robotics as a foundation-model problem rather than a hardware problem, an inversion of the hardware-first posture of Tesla Optimus and Figure. The profile raised the question of whether model-first robotics AI can compete with vertical hardware and software stacks, and suggested the company could become "the Anthropic/OpenAI of embodied AI" if its foundation-model-first approach succeeds (Source: TechCrunch / A peek inside Physical Intelligence, 2026-01-30).

The company frames the central technical challenge of robotics as generalization rather than agility or dexterity — the ability to perform even a simple task in a new setting or with new objects. It contrasts this with most commercial robots, which it says operate in tightly controlled environments such as factories or warehouses where objects and their locations are predetermined, and argues that robots in homes, stores, offices, and hospitals require stronger generalization (Source: https://www.pi.website/blog/pi05). In a typical deployment described by the analyst site Sacra, a developer streams RGB-D camera images from a robot to Physical Intelligence's runtime; the system tokenizes the visual stream together with the robot's movement history and feeds it to a transformer model of roughly 3 billion to 5 billion parameters, after which a hardware abstraction layer converts the model's output tokens into robot-specific joint commands within force and speed limits (Source: https://www.therobotreport.com/physical-intelligence-raises-600m-advance-robot-foundation-models/).

Models and research

Physical Intelligence's models are vision-language-action (VLA) models — derived from general vision-language models and trained to output robot actions alongside text and image tokens.

π0 (pi-zero)

π0 is the company's first generalist policy, described as a prototype model that combines large-scale multi-task and multi-robot data collection with a network architecture intended to produce a dexterous generalist robot policy. The company described π0 as "a small early step toward developing truly general-purpose robot models" and published an accompanying research paper (Source: https://www.therobotreport.com/physical-intelligence-raises-400m-for-foundation-models-for-robotics/). Early demonstrations included folding laundry, assembling boxes, scooping coffee beans, cleaning a table, and dynamically placing objects into containers (Source: https://www.therobotreport.com/physical-intelligence-open-sources-pi0-robotics-foundation-model/).

In February 2025 the company released the code and weights for π0 as part of an experimental repository called openpi on GitHub, along with checkpoints for tasks on the ALOHA and DROID robot platforms, example inference code, and fine-tuning code. The company said that between 1 and 20 hours of data was sufficient to tune π0 to a variety of tasks in its own experiments. Hugging Face prepared a PyTorch port of openpi for developers who prefer PyTorch over JAX. The release also included a π0-FAST variant that uses the FAST tokenizer to enable control through autoregressive discretization, which the company said offered somewhat better language-following performance at roughly four to five times higher inference cost (Source: https://www.therobotreport.com/physical-intelligence-open-sources-pi0-robotics-foundation-model/). In releasing openpi, the company wrote that it hoped open-source robotic foundation models would lead to new applications "in the same way that effective open-source language models (LLMs) and vision-language models (VLMs) have led to a Cambrian explosion" of applications (Source: https://www.therobotreport.com/physical-intelligence-open-sources-pi0-robotics-foundation-model/).

π0.5

π0.5, announced on April 22, 2025, is a model based on π0 that the company describes as exhibiting generalization to entirely new environments, such as cleaning a kitchen or bedroom in a home not seen in the training data. The company states that the main principle behind π0.5 is co-training on heterogeneous data — training the VLA on a mixture of robotic demonstrations, high-level subtask labels, verbal instructions, object detection, image captioning, visual question answering, and multimodal web data — to support generalization at the physical, visual, and semantic levels (Source: https://www.pi.website/blog/pi05). At inference, π0.5 first produces a high-level action expressed in language and then selects low-level motor commands, using discrete autoregressive token decoding for the high-level step and continuous flow matching through a roughly 300-million-parameter action expert for the low-level commands, in what the company likens to a chain-of-thought process (Source: https://www.pi.website/blog/pi05).

In ablation experiments reported by the company, the full π0.5 training mixture reached an 86% in-distribution language-following rate and an 83% in-distribution success rate, and 94% on both out-of-distribution follow and success rates; variants that excluded web data or data from other robots performed worse, with the largest out-of-distribution drops when multi-environment and cross-embodiment robot data were removed (Source: https://www.pi.website/blog/pi05). The company also reported a scaling study finding that generalization performance increased with the number of distinct training environments and approached a baseline trained directly on the test environment after about 100 training environments (Source: https://www.pi.website/blog/pi05). π0 and π0.5 were subsequently ported to PyTorch in the LeRobot library and made available on Hugging Face (Source: https://www.pi.website/blog/pi05).

π*0.6 and RECAP

In November 2025 the company announced version 0.6 of its VLA model, described as π*0.6. The model was trained with an approach the company calls RECAP (RL with Experience and Corrections via Advantage-conditioned Policies), which trains a robot by demonstration, coaches it through corrections, and improves it from autonomous experience. Physical Intelligence said this approach roughly doubled throughput on tasks such as inserting a filter into an espresso machine, folding previously unseen laundry, and assembling a cardboard box, decreased failure rates over hours of operation, and outperformed imitation learning alone (Source: https://www.therobotreport.com/physical-intelligence-raises-600m-advance-robot-foundation-models/).

Funding and valuation

Physical Intelligence raised $70 million in seed financing in 2024 (Source: https://www.therobotreport.com/physical-intelligence-raises-400m-for-foundation-models-for-robotics/). In November 2024 it announced a $400 million round at a $2.4 billion valuation, led by Jeff Bezos, Thrive Capital, and Lux Capital (Source: https://www.therobotreport.com/physical-intelligence-raises-400m-for-foundation-models-for-robotics/). In November 2025 it raised a $600 million Series B, led by CapitalG (Alphabet's growth fund) and Lux Capital, with participation from Bond, Redpoint, and Sequoia Capital; the company was valued at about $5.6 billion in that round according to Bloomberg, bringing total funding to about $1.1 billion (Source: https://www.therobotreport.com/physical-intelligence-raises-600m-advance-robot-foundation-models/). Earlier and returning investors named by the company and reporting include OpenAI, Khosla Ventures, Sequoia Capital, T. Rowe Price, and Redpoint Ventures (Source: https://www.therobotreport.com/physical-intelligence-raises-400m-for-foundation-models-for-robotics/) (Source: https://www.therobotreport.com/physical-intelligence-raises-600m-advance-robot-foundation-models/). The company has said it plans to use the Series B financing to collect more data, form strategic partnerships, and grow its team (Source: https://www.therobotreport.com/physical-intelligence-raises-600m-advance-robot-foundation-models/).

CapitalG partners Jill Greenberg Chase and Manmeet Gujral wrote that what they consider distinctive about the company's approach is "universally embodied AI": building "a single generalist intelligence that manifests in any physical form to solve any real-world problem" (Source: https://www.therobotreport.com/physical-intelligence-raises-600m-advance-robot-foundation-models/).

People

Physical Intelligence was founded in 2024 by a group that includes Karol Hausman (co-founder and chief executive), Sergey Levine (co-founder and an associate professor at the University of California, Berkeley), Chelsea Finn, Brian Ichter, Quan Vuong, Adnan Esmail, and Lachy Groom (Source: https://www.therobotreport.com/physical-intelligence-raises-400m-for-foundation-models-for-robotics/) (Source: https://www.therobotreport.com/physical-intelligence-open-sources-pi0-robotics-foundation-model/). Levine delivered a keynote on robotic foundation models at the RoboBusiness conference in 2024 (Source: https://www.therobotreport.com/physical-intelligence-raises-400m-for-foundation-models-for-robotics/).

Partnerships

Physical Intelligence has said it is pursuing partnerships to collect data and pretrain its models, and has expressed particular interest in collaborating with companies scaling up data collection with robots deployed for real-world applications (Source: https://www.pi.website/blog/pi05). The Chinese embodied-AI company AgiBot has announced a partnership with Physical Intelligence on embodied intelligence (Source: https://www.therobotreport.com/physical-intelligence-raises-600m-advance-robot-foundation-models/).

Anthropic held acquisition discussions with Physical Intelligence in spring 2026; the talks, and Anthropic's denial that any deal was in progress, became public on July 21, 2026 (Source: techcrunch.com).

Relationships

• deploys-in: Robotics (model layer; multiple embodiments).
• related: The Industrial Explosion (Davidson, Hadshar), Figure AI, planned Clone Robotics, Amazon (humanoid robots), Helix (Figure AI Vision-Language-Action model).

Sources

• TechCrunch / A peek inside Physical Intelligence (2026-01-30)
• The Robot Report / Physical Intelligence raises $600M to advance robot foundation models (2025-11-25) — https://www.therobotreport.com/physical-intelligence-raises-600m-advance-robot-foundation-models/
• The Robot Report / Physical Intelligence open-sources Pi0 robotics foundation model (2025-02-07) — https://www.therobotreport.com/physical-intelligence-open-sources-pi0-robotics-foundation-model/
• The Robot Report / Physical Intelligence raises $400M for foundation models for robotics (2024-11-05) — https://www.therobotreport.com/physical-intelligence-raises-400m-for-foundation-models-for-robotics/
• Physical Intelligence / π0.5: a VLA with Open-World Generalization (2025-04-22) — https://www.pi.website/blog/pi05