Physical Intelligence, a two-year-old robotics startup based in San Francisco, is pursuing a research-first approach to general-purpose robotic intelligence, prioritising large-scale data collection and model training over near-term commercialisation, while raising more than $1bn in funding and operating at a reported $5.6bn valuation.
Inside The Company’s Research Operations
Physical Intelligence operates from a sparse headquarters in San Francisco with no visible branding or reception area. Inside, long shared tables are used both for meals and for robotics experiments, with robotic arms set up to perform everyday tasks such as folding clothes, turning garments inside out, and peeling vegetables. These tasks are not product demos but part of a continuous training and evaluation loop.
According to Sergey Levine, a co-founder of Physical Intelligence and an associate professor at University of California Berkeley, data collected from robot stations in warehouses, homes, and test kitchens is used to train general-purpose robotic foundation models. Updated models are then re-evaluated on the same stations. The aim is to teach robots fundamental physical skills that can transfer across tasks and environments.
The company also runs test kitchens using standard commercial equipment, including espresso machines, to expose robots to varied physical settings. These environments are designed to generate training data rather than serve operational needs for staff.
Focus On Software Over Hardware
The robotic arms used in testing are commercially available models priced at around $3,500, which Levine says includes a large vendor markup. He estimates the material cost would fall below $1,000 if manufactured internally. The company’s strategy is to rely on relatively basic hardware and focus resources on intelligence and learning systems, based on the view that stronger models can compensate for limited physical capabilities.
Founding Team And Origins
Physical Intelligence was founded by researchers including Levine, Chelsea Finn, who runs a robotics lab at Stanford University, and Karol Hausman, who is affiliated with Google DeepMind.
The company’s chief executive is Lachy Groom, who previously worked at Stripe and spent several years as an angel investor backing companies such as Figma, Notion, Ramp, and Lattice. Groom says he was actively searching for a company to start or join after leaving Stripe and was drawn to the academic work of Levine and Finn before becoming involved.
Funding And Spending Priorities
Physical Intelligence has raised more than $1bn from investors including Khosla Ventures, Sequoia Capital, and Thrive Capital. Groom says most of the company’s spending goes toward compute rather than headcount or hardware. He adds that while the company does not burn cash aggressively, it would raise additional funding under favourable terms, citing ongoing demand for more computing resources.
Groom says the company does not provide investors with a timeline for commercial revenue, a position he acknowledges is unusual. He says investors have accepted the absence of defined monetisation plans, in part due to the company’s current capital position.
Strategy Based On Transferable Learning
Another co-founder, Quan Vuong, formerly of Google DeepMind, says the company’s strategy centres on cross-embodiment learning. The goal is to ensure that models trained on one robotic platform can transfer knowledge to new hardware without restarting data collection. Vuong says this reduces the marginal cost of enabling autonomy on new robotic systems.
Physical Intelligence is working with a limited number of partners across logistics, grocery, and food production to test whether its models can handle real-world automation tasks. Vuong says some deployments already meet operational requirements, depending on the task.
Competitive Landscape In Robotics
The company is operating in a crowded field focused on general-purpose robotic intelligence. Skild AI, founded in 2023 and based in Pittsburgh, raised $1.4bn this month at a $14bn valuation. Skild AI has taken a more commercial approach, deploying its Skild Brain system across security, warehousing, and manufacturing and reporting $30m in revenue within months of launch.
Skild AI has publicly criticised competing approaches, arguing that many robotics foundation models rely too heavily on internet-trained vision-language systems rather than physics-based simulation and real-world robotics data.
Workforce And Operational Challenges
Physical Intelligence employs about 80 people and plans to grow gradually. Groom says hardware remains the most difficult aspect of the business, citing long lead times, frequent breakages, and safety constraints that slow testing and iteration compared with software development.
The company maintains an internal roadmap projecting capabilities over five to ten years. Groom says many of those milestones were reached within the first 18 months of operation.
Featured image credits: MicroMain CMMS
For more stories like it, click the +Follow button at the top of this page to follow us.