- Physical Intelligence’s π0.7 model can direct robots to perform tasks they were never trained on—by recombining learned skills in ways the company’s own researchers didn’t anticipate.
- The model folded laundry on a new robot platform using zero laundry-folding data, and operated kitchen appliances it had never seen before with just language coaching.
- The San Francisco startup, valued at $11 billion with Jeff Bezos among its backers, published the research Thursday alongside a video demonstration of the model’s emergent capabilities.
Physical Intelligence, the two-year-old San Francisco robotics startup that has quietly become one of the most closely watched AI companies in the Bay Area, published new research Thursday showing its latest model can handle tasks it was never explicitly trained on. The company’s own researchers say the results caught them off guard.
The model, called π0.7, demonstrates what the company describes as “compositional generalization”—the ability to combine skills learned in different contexts to solve entirely new problems. Think of it like a chef who trained in French cuisine, Japanese knife work, and Mexican grilling, then walked into a Thai kitchen and figured out pad thai from scratch. Until now, robot training has been closer to rote memorization: collect data on one task, train a specialist model, throw it away, repeat.
According to TechCrunch, the researchers first noticed the emergent behavior when they tasked the model with operating kitchen appliances it had never encountered. The robot received step-by-step language coaching—essentially the same instructions you’d give a person using an air fryer for the first time—and made a “reasonable attempt” at cooking a sweet potato. Not a perfect attempt, but the fact that it worked at all with zero appliance-specific training data was the surprise.
How π0.7’s Compositional Generalization Actually Works
The key to π0.7’s generalization isn’t just more data—it’s how the model is prompted during training. Physical Intelligence’s approach uses what it calls “diverse multimodal prompts”: not just language instructions telling the robot what to do, but also metadata specifying how fast or how well, visual subgoal images showing what the end result should look like, and control modality labels for different robot types.
This prompting framework lets the model absorb data from wildly different sources—multiple robot platforms, human demonstration videos, and even lower-quality autonomous episodes—without getting confused about what good performance looks like. At test time, the model accepts plain language instructions and can even use synthetically generated visual subgoals produced by a lightweight world model.
The results are striking. In one demonstration, π0.7 folded laundry on a robot platform for which there is no laundry folding training data whatsoever. In another, it picked up and used kitchen appliances by combining manipulation skills it learned separately across different tasks. The model also generalizes across robot platforms more effectively than Physical Intelligence’s prior models, matching the performance of fine-tuned specialist systems straight out of the box.
If these findings hold up to outside scrutiny—and robotics has a long history of impressive demos that don’t survive real-world deployment—they suggest robot AI may be approaching the kind of compounding capability gains that large language models experienced around GPT-3. Where LLMs can compose concepts from training data in new ways (translate to French, then format as JSON), vision-language-action models have until now lacked that same compositional spark.
Physical Intelligence is reportedly in talks for a $1 billion funding round that would value the company at north of $11 billion, doubling its previous valuation. Jeff Bezos is among its investors, and the company has raised more than $1 billion to date. Some of its founders got their start at Google DeepMind.
The robot that learned to fold laundry without ever being shown how to fold laundry is, for now, the most compelling evidence that general-purpose robot brains aren’t just a pitch deck promise.