Robotic Foundation Models are reshaping everyday automation — from zucchini peeling to warehouse brains, right now.
Silicon Valley’s newest robotic buzz centers on Robotic Foundation Models: general-purpose brains for all kinds of robots. Physical Intelligence is building those brains, training models on real-world tasks from kitchens to warehouses. The company has raised over $1 billion and is valued at $5.6 billion, backing a long-term play rather than near-term commercialization. For context on how AI debate and layered reasoning improves outcomes, see my write-up on Society of Thought — the ideas overlap in surprising ways.
As VP Emerging Technologies at Ericsson I obsess over where compute meets the messy physical world. I once tried teaching a home robot to make espresso — it frothed the milk onto my keyboard. That failure taught me what I already knew: soft intelligence often beats flashy hardware. My background in IoT and robotics makes Physical Intelligence’s focus on broad, transferable data feel very familiar — and oddly comforting.
Robotic Foundation Models
Physical Intelligence is literally running robot experiments in a concrete box in San Francisco. The scene is humble: $3,500 off-the-shelf arms (“an enormous markup,” the team says) that would cost under $1,000 in parts if made in-house. The point is obvious — smart models compensate for cheap hardware. Co-founder Sergey Levine summed it up plainly: “Think of it like ChatGPT, but for robots.” That quote, and hands-on examples like a zucchini-peeler and a pants-folder, explain both the ambition and the grind.
How they collect data
Data comes from a continuous loop. Robots at test stations in homes, warehouses and a test kitchen generate real physical interactions. That data trains general-purpose models which are then returned to stations for evaluation. The company has raised over $1 billion and is valued at $5.6 billion, so they can keep throwing compute at the problem — and they do. Co-founder Lachy Groom noted most spending goes to compute; “There’s always more compute you can throw at the problem,” he said.
Cross-embodiment and transfer
Physical Intelligence bets on cross-embodiment learning. Quan Vuong explains the marginal cost of onboarding a new robot platform drops dramatically because a trained model carries transferable skills. That means an arm that learns peeling zucchini should generalize to peeling apples or potatoes it never saw. This “any platform, any task” philosophy aims to reduce deployment friction for partners in logistics, grocery and manufacturing.
Competition and strategy
The company is not alone. Skild AI raised $1.4 billion this month at a $14 billion valuation and reported $30 million in revenue from rapid deployments. Skild argues commercial deployments create a data flywheel; Physical Intelligence resists near-term commercialization to prioritize research purity. Which approach wins will take years. For now, the stakes are empirical: models trained on real interactions may achieve the “physical common sense” critics say vision-language models lack.
Why it matters
If Robotic Foundation Models succeed, the industry gets reusable, transferable robot brains. The surface area for success is large: small automation wins today pay the bills while research builds toward broader capabilities. Physical Intelligence’s mix of cheap hardware, heavy compute, and a two-year sprint backed by elite researchers could reshape how we automate kitchens, warehouses and factories — one peeled zucchini at a time. See the full TechCrunch report at TechCrunch.
Robotic Foundation Models Business Idea
Product: A SaaS+Edge offering called “PlatformBrain” that packages pretrained Robotic Foundation Models tuned for verticals (grocery, logistics, light manufacturing, food prep). PlatformBrain supplies model access, data connectors, simulation-based fine-tuning, and a hardware-agnostic SDK to onboard new robot platforms in weeks.
Target market: Mid-size warehouses, automated dark kitchens, grocery fulfillment centers, and food manufacturers who need rapid deployment without building in-house robotics teams.
Revenue model: Subscription tiers ($5k–$50k/month) for model access and cloud training; per-device licensing for edge inference; professional services for integration and custom fine-tuning; revenue share on automation efficiency gains for large partners.
Why now: Robotic Foundation Models reduce onboarding costs and speed time-to-value. With suppliers offering <$1k hardware and investors funding compute-heavy research, a product that operationalizes those models fills a clear commercialization gap. Early revenue from deployments funds domain-specific data collection, accelerating model improvement and creating a defensible data moat.
Beyond the Prototype
Robotic Foundation Models promise durable automation: reusable skills that travel across arms, kitchens and warehouses. Physical Intelligence is placing a long bet on research-first progress. That purity could produce the robust, general physical intelligence industry needs. What small task would you hand over first to a robot trained by these models?
FAQ
Q: What are Robotic Foundation Models?
A: General-purpose models trained on diverse real-world robot interactions. They enable transfer of skills across platforms and tasks, reducing onboarding cost and time for new hardware.
Q: How much has Physical Intelligence raised and what does that fund?
A: The company raised over $1 billion and is valued at $5.6 billion. Most spending goes to compute for training models and running continuous data loops across test stations.
Q: How do these models differ from vision-language approaches?
A: Robotic Foundation Models prioritize physics and real interaction data. Critics say vision-language models rely on internet pretraining; physical models use hands-on data and simulation to learn “physical common sense.”