The turning point for robots has arrived? This American company claims its new model can “enable robots to perform tasks they were never trained for.”

```

The field of robotics AI may be ushering in a capability leap similar to that of large language models.

San Francisco-based robotics startup Physical Intelligence released its latest research on Thursday, claiming that its new model π0.7 can direct robots to complete tasks it has never specifically trained for—a capability that even surprised the company’s own researchers.

The company’s co-founder, UC Berkeley professor Sergey Levine, stated that this marks a shift in robotics AI from "rote memorization" to "analogy and inference," with its rate of capability improvement outpacing the linear growth of training data.

If this breakthrough is externally validated, it could have a profound impact on the commercialization trajectory of the robotics industry—robots could be deployed in new environments and optimized in real time without the need for additional data collection or retraining of models. Meanwhile, according to reports, Physical Intelligence is in talks for a new round of financing that could nearly double its valuation from $5.6 billion to $11 billion.

Core Breakthrough: From "Task-Specific Memory" to "Combinatorial Generalization"

Physical Intelligence was founded just two years ago. The core capability exhibited by the newly released π0.7 model is termed “compositional generalization” by researchers—the ability to combine skills acquired in different contexts to solve previously unseen problems.

This is fundamentally different from previous mainstream paradigms of robot training. The standard practice was essentially "rote memorization": collecting data for each specific task, training specialized models, and repeating this process for each new task. π0.7 breaks this cycle.

Levine compared this shift to a leap seen in large language models: "Once that critical point is crossed, you go from only being able to perform data-supported tasks to recombining skills in new ways—the speed of capability improvement then outpaces the linear increase in data. We’ve previously observed this more favorable scaling property in language and vision domains."

Key Demonstration: Air Fryer Experiment Reveals "Emergent Knowledge"

The most compelling demonstration in this research involved an air fryer the model had almost never encountered during training. The research team later found out there were only two related entries in the entire training dataset: one where a different robot pushed the fryer closed, and another from an open-source dataset of a robot, according to instructions, placing a plastic bottle inside it.

Nevertheless, π0.7 integrated these fragmented pieces of information with broader web pre-training data to develop a functional understanding of the appliance. With zero prompting, the model attempted to cook sweet potatoes using the air fryer, achieving basically acceptable results; with step-by-step language instructions, it completed the task successfully.

Physical Intelligence researcher and Stanford computer science Ph.D. student Lucy Shi described the dramatic shift in an early experiment: Success rates were only 5% at first, but after about half an hour optimizing how the task was described, success jumped to 95%. "Sometimes the failure doesn’t lie with the robot or the model, but with ourselves—not having engineered the prompts well enough," she said.

Research scientist Ashwin Balakrishna added that in the past, he could always predict a model’s bounds based on its training data. "But in the past few months, I was genuinely surprised for the first time. I casually bought a set of gears and asked the robot if it could turn them, and it just did it."

Limitations: Researchers Set Boundaries Proactively

The research team is candid about the model’s limitations. π0.7 still can’t autonomously complete complex multi-step tasks starting from a single high-level instruction. "You can’t just say, 'make me a piece of toast'," Levine noted, "but if you guide it step by step—'for the toaster, open this part, press that button, do this'—it usually does it quite well."

Additionally, the lack of standardized benchmarks in robotics makes external validation quite challenging. Physical Intelligence chose to compare π0.7 to their previous specialized models, and results show this generalist model matches specialist-level performance in complex tasks like making coffee, folding clothes, and assembling boxes.

The paper itself is cautious in tone, describing π0.7 as showing "early signs" and "preliminary demonstrations" of generalization. When directly asked when systems based on this research could be deployed in the real world, Levine declined to make predictions: “I believe there’s plenty of reason for optimism, and progress is faster than I expected two years ago. But that’s a hard question for me to answer.”

Capital Bets: Valuation May Double to $11 Billion

Physical Intelligence has raised over $1 billion to date, with a latest valuation of $5.6 billion. According to reports, the company is currently in talks for a new round of financing that may push its valuation close to $11 billion.

Investor enthusiasm for the company is in large part due to co-founder Lachy Groom. Groom was one of Silicon Valley’s most recognized angel investors, having backed well-known companies like Figma, Notion, and Ramp. Before co-founding Physical Intelligence, he considered it the company he had long been searching for. This background helped the startup attract institutional capital, even though the company consistently refuses to provide investors with a commercialization timeline.

Levine, addressing potential criticism from outsiders, anticipated the main objection: "For any robotics generalization demo, the criticism is always—'the tasks are too boring, and the robot isn’t doing backflips.'" He countered that truly generalizable robotic systems will always appear less impressive than carefully choreographed stunts, but are far more valuable in practice.

Risk Reminder and DisclaimerThe market has risks; investment requires caution. This article does not constitute personal investment advice and does not take into account the special investment objectives, financial situation, or needs of any particular user. Users should consider whether any opinions, viewpoints, or conclusions in this article fit their own circumstances. Investing accordingly is at your own risk. ```