Physical Intelligence pitches its π models as a shared robot brain, with Weave and Ultra as first proof at 96.4% autonomy

Physical Intelligence, the US start-up that has raised one of the largest war chests in robot software, set out its commercial thesis in public on February 24, 2026. In a post titled “The Physical Intelligence Layer”, the company argued that its general-purpose robot models should become a reusable layer that other companies build robot products on, in the same way that software developers now call large language models through an API instead of training their own. It backed the argument with field results from two robot makers, Weave and Ultra, both running its π0.6 model.

The headline numbers came from those two partners. At a customer site of Ultra, which builds warehouse robots, π0.6 packed e-commerce orders for a full shift at 96.4% autonomy, Physical Intelligence said. At Weave, which builds home robots, π0.6 lifted the share of time the robot folded laundry without help compared with the earlier π0.5, and adding Weave’s own data into pre-training cut missed grasp sequences by 42% and human interventions per full laundry load by 50%, according to the company. Physical Intelligence did not disclose prices, API terms, availability dates or whether model weights would be shared with partners.

What was announced

The post named four models as candidates for the layer: π0, π0.5, π0.6 and π*0.6. Physical Intelligence said it works with companies that specialise in deploying robots, both to show that its models perform in real settings and to stress-test how general they are across tasks and environments. For Ultra, the company said π0.6 raised the packing success rate over π0.5 and that adding Ultra pre-training data raised order-packing throughput, measured in items per hour; the absolute throughput figures appeared only in a chart. Physical Intelligence added that it is working with “a number of other partners” it plans to name later.

The two partners sit at opposite ends of the market. Weave describes itself as a builder of robots for the home; its website says its stationary Isaac 0 laundry-folding robot ships across California and that the Isaac 1 home robot is due to start shipping in fall 2026. Ultra builds industrial AI robots that slot into existing workstations. Its OP1 packs, sorts and kits items for up to 24 hours a day on a 5 by 5 foot footprint with a payload of up to 10 lb and plugs into a standard 120V outlet, according to Ultra, which says the robot is built in New York and offers customers a three-month trial.

The company behind the models

Physical Intelligence is led by co-founder and chief executive Karol Hausman; its founding team includes Lachy Groom, Sergey Levine, Chelsea Finn, Brian Ichter, Quan Vuong and Adnan Esmail, according to CapitalG, the Alphabet growth fund that led its Series B announced on November 20, 2025. Investor Outset Capital said that round raised $600 million at a $5.6 billion valuation, with Lux Capital, Thrive Capital, Jeff Bezos, Index Ventures and T. Rowe Price taking part. CapitalG described π0.6 as a model that learns through instruction, coaching and then autonomous practice, and said success rates exceed 90% on tasks such as making espresso and folding laundry.

Rivals take different routes

The pitch puts Physical Intelligence in direct competition with other suppliers of robot brains, each with a different business model. Skild AI raised $1.4 billion at a valuation above $14 billion in January 2026 on the promise of one model that can run any robot body. Rhoda AI left stealth on March 10 with a $450 million Series A for a model pre-trained on internet video for factory robots. Figure keeps its Helix model in-house and released Helix 02 for whole-body control of its own humanoid in January. NVIDIA, for its part, released GR00T N1.6 as an open model at CES, and Google DeepMind supplies its Gemini Robotics models to hardware partners such as Agile Robots, which signed a research pact in March.

How the layer is meant to work

The models in question are vision-language-action models: networks that take camera images and a plain-language instruction and output motor commands. A partner collects demonstrations on its own robot and task, fine-tunes the shared model, and ships the result. The Weave and Ultra figures show a second step that matters commercially. The biggest gains came when partner data was folded back into Physical Intelligence’s pre-training, which suggests that every deployment can improve the base model for everyone, and that the supplier of the base model gains from each customer’s data.

A week later, on March 3, the company showed how it intends to stretch the layer to longer jobs. Its Multi-Scale Embodied Memory system, described in an arXiv paper with co-authors from Stanford University, UC Berkeley and MIT, adds a compressed video memory for recent detail and text summaries for longer-range events to π0.6, which uses Google’s Gemma 3 4B model as its language backbone. The paper reports tasks such as kitchen clean-up and making a grilled cheese sandwich with memory spanning up to 15 minutes.

What it means for the industry

If the layer works as described, the entry ticket for a robot start-up shrinks from building a full AI stack to building hardware, collecting task data and integrating a model. That would shift value toward whoever owns the most capable base model and the widest pool of deployment data. The open questions are large. Physical Intelligence has published no commercial terms, the results come from two partners and are self-reported, and 96.4% autonomy over a shift still implies human help for part of the time. Data rights, the cost of inference on the robot and liability when a licensed model fails remain undisclosed.