Physical Intelligence, a hot robotics startup, says its new robot brain can figure out tasks it was never taught

Physical Intelligence, a two-year-old robotics startup based in San Francisco, has recently unveiled groundbreaking research that highlights its latest model's ability to direct robots in performing tasks without having been explicitly trained for them. This unexpected capability has taken the company's researchers by surprise and marks an important step toward achieving a general-purpose robotic brain. The new model, named π0.7, is designed to tackle unfamiliar tasks by utilizing plain language coaching from humans. This innovative approach could signal a turning point in robotic AI technology, reminiscent of the advancements witnessed in large language models, where capabilities expand rapidly beyond initial expectations. At the core of this research is the concept of compositional generalization, which allows the model to combine previously learned skills from various contexts to solve new problems. Traditionally, robot training has relied heavily on rote memorization, requiring specific data collection and model training for each individual task. However, Physical Intelligence asserts that π0.7 breaks away from this conventional pattern. Sergey Levine, co-founder of Physical Intelligence and an AI for robotics professor at UC Berkeley, explains that the model's ability to remix learned skills leads to a more than linear increase in capabilities with additional data. This shift could redefine how robots are trained and deployed in various environments. One of the most impressive demonstrations involved an air fryer, a device the model had never encountered during training. The researchers found only two relevant instances within the training dataset, yet the model managed to synthesize this limited information with broader web data to operate the appliance effectively. Without any coaching, it attempted to cook a sweet potato, and with step-by-step verbal instructions, it succeeded. This coaching capability is significant, as it implies robots could be quickly adapted to new environments without the need for extensive retraining. Despite its achievements, the researchers remain cautious about the model's limitations. They acknowledge that failures can stem from their own prompt-engineering skills, which significantly impact the model's performance. Levine emphasizes that while the model can successfully follow detailed instructions, it is not yet capable of executing complex tasks autonomously from a single command. Furthermore, standardized benchmarks for robotics are lacking, making external validation of their claims challenging. In comparing π0.7 to previous specialized models, the team found that it matched or exceeded performance across various tasks, including making coffee and folding laundry. The most remarkable aspect of this research is the researchers' own surprise at the model's capabilities, which suggests that even those intimately familiar with the training data can be astonished by the results. While the potential for generalization in robotics is exciting, critics point out that robots have not had the vast resources available to language models. Levine acknowledges this but argues that the distinction between impressive demonstrations and practical generalization is crucial. The researchers are careful to describe π0.7 as showing initial signs of generalization, emphasizing that these results are still in the research phase. Physical Intelligence, having raised over $1 billion and recently valued at $5.6 billion, continues to draw interest from investors, with discussions underway for further funding that could elevate its valuation to $11 billion. However, the company has not provided a timeline for when these advancements might be ready for real-world application.