New research from Physical Intelligence, a robotics startup based in San Francisco, reveals its latest model, π0.7, can enable robots to accomplish tasks without prior specific training. This capability has surprised the company’s researchers, who believe it signifies a potential turning point in robotic AI, akin to advancements seen in large language models.
The model showcases an ability known as compositional generalization, allowing it to integrate skills from various contexts to tackle unfamiliar problems. Traditionally, robots have been trained through rote memorization, requiring extensive data collection for each unique task. However, π0.7 appears to deviate from this method, presenting a more efficient learning paradigm.
One notable demonstration involved an air fryer, which the model had not encountered during training. The research team identified only two instances related to the appliance in their dataset. Nonetheless, π0.7 successfully combined this limited information with general pretraining data to understand the air fryer's functionality.
Sergey Levine, co-founder of Physical Intelligence, emphasized that once the model surpasses basic task completion, its capabilities can expand more significantly than expected with additional data, similar to trends in language and vision AI.