Physical Intelligence Unveils π0.7, a Steerable Robot Foundation Model That Learns New Tasks From Language Coaching

Overview

Physical Intelligence, the San Francisco robotics startup founded by former Google DeepMind researchers and Stripe veteran Lachy Groom, released a new generalist robot model called π0.7 on April 16, along with a research paper detailing what the company describes as early signs of compositional generalization in robotic manipulation. The model, described by TechCrunch as a step toward a general-purpose robot brain, can be pointed at an unfamiliar task, coached through it in plain language, and — in the company’s demonstrations — actually pull it off.

The release lands as Physical Intelligence is reportedly in talks to raise roughly 1 billion dollars at a valuation north of 11 billion dollars, roughly doubling the 5.6 billion dollar mark set four months earlier. TechCrunch reports the new round would push total funding past the 2 billion dollar threshold in a two-year-old, roughly 80-person company.

What the Model Does

The π0.7 paper on arXiv frames the model as a “steerable generalist robotic foundation model with emergent capabilities.” Rather than the standard robotics pipeline — collect data on one task, train a specialist, repeat — π0.7 is a single model trained on data from many robots, human videos, and autonomous episodes, including failures. The paper claims zero-shot cross-embodiment generalization, meaning the same policy can drive different robot bodies on tasks it was not specifically trained for, and argues this is enabled by diverse multimodal prompting during training.

On the Physical Intelligence side of the benchmarking, the single generalist model reportedly matches or exceeds the performance of the company’s earlier task-specific RL-tuned specialists on laundry folding, espresso making, and box assembly. TechCrunch reports that when the team pointed π0.7 at an air fryer and asked it to cook a sweet potato, the robot initially succeeded about 5 percent of the time. After roughly half an hour of prompt refinement — adjusting how the task was verbally explained to the model — the success rate jumped to 95 percent.

Research scientist Ashwin Balakrishna told TechCrunch that a separate test with a randomly purchased gear set illustrated the pattern: “I just bought a gear set randomly and asked the robot, ‘Hey, can you rotate this gear?’ And it just worked.” Co-founder Sergey Levine framed the broader claim this way in the same article: “Once it crosses that threshold where it goes from only doing exactly the stuff that you collect the data for to actually remixing things in new ways, the capabilities are going up more than linearly with the amount of data.”

What We Know

• The release was published on April 16, 2026, with a companion arXiv paper submitted the same day, according to the paper’s arXiv listing.
• The model is trained on data from multiple robot embodiments, human video, and autonomous episodes — including suboptimal runs and failures — per the arXiv abstract.
• TechCrunch reports the air fryer sweet potato task improved from 5 percent to 95 percent success after about 30 minutes of prompt refinement, according to Physical Intelligence’s demonstration described by TechCrunch.
• The company has raised more than 1 billion dollars to date and was most recently valued at 5.6 billion dollars, with a new round targeting an 11 billion dollar valuation under discussion, as reported by TechCrunch.
• Investors in the prospective round reportedly include Founders Fund and Lightspeed Venture Partners, alongside returning backers Thrive Capital and Lux Capital, according to TechCrunch.

What We Don’t Know

The company’s own benchmarks are the main source for π0.7’s performance figures, and TechCrunch notes that standardized robotics benchmarks comparable to language model leaderboards do not yet exist, limiting external validation. The air fryer demonstration required a human operator to iterate on prompt wording for about half an hour before reaching 95 percent success, meaning the robot’s out-of-the-box performance on a novel appliance was 5 percent — a gap the article acknowledges when it notes the model “cannot execute complex multi-step tasks autonomously from single commands.”

Physical Intelligence has not disclosed a commercialization timeline. Co-founder Lachy Groom told TechCrunch in March that “there’s no limit to how much money we can really put to work,” and the company has not announced paying customers or hardware partners for π0.7 specifically. Independent reproduction of the cross-embodiment claims will depend on whether the model weights, evaluation scripts, or underlying data are released in a form outside researchers can run.

Analysis

The π0.7 release fits a pattern visible across robotics foundation models this year: vendors increasingly argue that language coaching and multimodal prompting can substitute for the expensive task-specific data collection that has gated deployment. Physical Intelligence’s framing — that success rates improve sharply once the model crosses a generalization threshold — echoes the talking points used by competitors racing to build what they variously call a “robot brain” or a “world model.”

What distinguishes the π0.7 announcement is the combination of a peer-reviewable arXiv paper, a named set of tasks (laundry folding, espresso making, box assembly, air fryer operation, gear rotation), and an explicit acknowledgement that out-of-the-box performance on a cold task can be as low as 5 percent before prompting. That honesty about the gap between demo reel and deployment is unusual in a sector where humanoid vendors routinely publish edited videos without disclosing success rates. Whether π0.7’s steerability holds up when evaluated by parties who did not write the prompts — and whether the 11 billion dollar funding round closes — will be the near-term tests of how much of the hype survives contact with customers.