AIWiki
Malaysia
Back to all articles
Applicationsphysical-intelligenceroboticsvision-language-action

Physical Intelligence

4 min readUpdated November 2026
Physical Intelligence
Type
Robotics AI research company
Founded
February 2024, San Francisco
Founders
Karol Hausman, Sergey Levine, Chelsea Finn, Brian Ichter, Lachy Groom, Adnan Esmail, Quan Vuong
Flagship models
pi0, pi0.5 (open weights via openpi)
Funding
$400M (2024); $600M at ~$5.6B (2025); ~$1B round reported 2026
Related
Vision-language-action models, Physical AI, Flow Matching
Physical Intelligence is an American robotics company founded in February 2024 and based in San Francisco, which builds foundation models for robot control — a single neural network, trained on data from many robots, that can be fine-tuned to perform many different physical tasks. Its models, beginning with pi0, aim to do for robotics what large language models did for text: one general system instead of a separately engineered solution for every task.[1]

History

The company was founded by researchers from Google DeepMind (Karol Hausman, Brian Ichter, Quan Vuong), Stanford (Chelsea Finn) and UC Berkeley (Sergey Levine), together with Adnan Esmail and entrepreneur Lachy Groom. It emerged from years of academic work on robot learning — algorithms that let machines learn manipulation skills from demonstration data rather than hand-coded trajectories.[1]

Physical Intelligence raised $70 million in seed funding in early 2024, followed by $400 million in November 2024 at a $2.4 billion valuation.[2] In 2025 it raised a further $600 million led by CapitalG at roughly $5.6 billion, with participation from Lux Capital, Thrive Capital, Jeff Bezos, Sequoia and T. Rowe Price; in 2026 Bloomberg reported the company was seeking $1 billion at around $11 billion.[1]

In October 2024 the company released pi0 and its technical paper. In February 2025 it open-sourced the code and weights as openpi, and in September 2025 released pi0.5, an upgraded model with better open-world generalisation; further versions have followed in the openpi repository.[3][4]

Key Concepts and Technology

pi0 is a vision-language-action (VLA) model: it combines a vision-language model that perceives images and interprets instructions with an action-expert flow-matching model that outputs continuous motor commands (joint positions, gripper states) at high frequency.[5] Flow matching, related to the technique behind diffusion models, lets the model sample smooth action trajectories conditioned on what it sees and what it was asked to do.

Training uses a large, heterogeneous mixture of data from many different robots — dual-arm arms, single-arm arms, mobile manipulators — which is what gives the policy transfer across hardware. pi0.5 adds co-training on heterogeneous tasks so the model can follow novel instructions in unfamiliar environments, the property the company calls open-world generalisation. Because the weights are open, researchers and companies can fine-tune pi0 on their own robots and tasks.[3]

Applications and Impact

Reported and demonstrated applications include folding laundry, clearing and bussing tables, assembling cardboard boxes, packing e-commerce orders and handling objects in laboratory and industrial settings. The broader impact is a shift in the robotics industry from task-specific programming toward generalist policies — where a deployer buys a model and gathers a small amount of demonstration data for the target site, instead of engineering a controller from scratch.

>See Also

🇲🇾Malaysian Context

🇲🇾 Relevance to Malaysia: Robotics sits close to Malaysia's economic priorities. Penang's semiconductor and electronics cluster, Malaysia's medical-device and electrical-and-electronics manufacturing, and the wider Industry 4.0 push under the MyDigital Blueprint all depend on automation, and generalist robot models lower the cost of automating high-mix, low-volume production lines where traditional industrial automation is too rigid. A Malaysian manufacturer could fine-tune an open model such as pi0 on its own demonstration data rather than commissioning a bespoke integrator solution — an approach suited to Malaysian SMEs' budgets. Universities and TVET institutions gain a teachable platform for robotics and physical AI, an area the National AI Office has flagged as strategically relevant, while SIRIM and MDEC-standard safety and data practices apply where robots handle sensitive processes. As with all VLA systems, deployment requires human oversight: these models are policies, not guarantees, and Malaysian factories must retain safety certification and human-in-the-loop controls.

References

  1. ↑Wikipedia. Physical Intelligence Inc. https://en.wikipedia.org/wiki/Physical_Intelligence_Inc.
  2. ↑The Robot Report (2024). Physical Intelligence raises $400M for foundation models for robotics. https://www.therobotreport.com/physical-intelligence-raises-400m-for-foundation-models-for-robotics/
  3. ↑Physical Intelligence (2025). Open Sourcing pi0. https://www.pi.website/blog/openpi
  4. ↑GitHub. Physical-Intelligence/openpi. https://github.com/Physical-Intelligence/openpi
  5. ↑Black et al. (2024). pi0: A Vision-Language-Action Flow Model for General Robot Control. arXiv:2410.24164. https://arxiv.org/abs/2410.24164
Back to all articles