New AI model enable robots to perform unseen tasks, hinting at a shift toward general-purpose robotic intelligence.
A US robotics startup says its latest AI model can guide robots to perform tasks they were never trained on.The system, developed by Physical Intelligence , called π0.7 , marks an early step toward a general-purpose robot brain that can handle unfamiliar tasks using plain-language guidance. According to the San-Francisco-based firm, the results were unexpected. If confirmed, they suggest robotic AI may be nearing a turning point, with capabilities advancing faster than anticipated.“In our experiments, we see π0.7 exhibiting the first signs of compositional generalization, recombining skills from various tasks to solve new problems, like using new kitchen appliances and even enabling a new robot to fold laundry for which there is no laundry folding data,” said the firm in a blog post.
Robot brain emergesThe new model, π0.7, is described as an early but meaningful step toward a general-purpose robot brain capable of handling unfamiliar tasks through plain-language guidance. According to researchers, it demonstrates a clear advance in generalization, performing a wide range of dexterous tasks at a level comparable to specialized systems while also executing tasks not included in its training data.Experiments show early signs of compositional generalization, where the model recombines learned skills to solve new problems. Examples include using unfamiliar kitchen appliances and enabling a robot to perform tasks like folding laundry without prior task-specific data. This marks a shift from traditional robot training, which relies on collecting data and building separate models for each task.
Unlike earlier vision-language-action systems, which struggled to combine skills in novel ways, π0.7 can apply existing capabilities across new contexts without additional fine-tuning. It also generalizes more effectively across different robots, environments, and tasks.These results suggest a transition from task-specific training toward more flexible, general-purpose systems, where capabilities scale more efficiently as models learn to reuse and recombine knowledge across domains.
Flexible AI systemsπ0.7’s broad generalization stems from how it is trained and prompted. Instead of relying on a single data source, the model is built on a mix of inputs, including multiple robot platforms, human demonstrations, and autonomously collected episodes. Rather than combining these datasets directly, the system is trained with rich, multimodal prompts that define not only the task but also execution details. These prompts can include text instructions, visual subgoals that specify object arrangements, and parameters such as task duration. This added context helps the model interpret diverse behaviors and strategies, enabling it to apply knowledge more flexibly.
During inference, the model can follow standard language instructions while also incorporating guidance, such as desired strategies or generated visual targets. This allows it to adapt in real time, improving performance without retraining. In testing, the system demonstrated the ability to infer how to use unfamiliar objects by combining limited prior examples with broader learned knowledge. With minimal guidance, it attempted new tasks, and with structured step-by-step instructions, performance improved significantly.
The approach highlights a shift toward interactive learning, where human feedback and prompt design play a critical role in outcomes. However, the system still requires detailed guidance for multi-step tasks and cannot autonomously execute complex instructions from a single command. Researchers also note the absence of standardized benchmarks, making independent validation difficult. The findings remain early-stage, but they point to more adaptable robotic systems that can extend capabilities beyond their original training.
AI Model AI Systems Artificial Intelligence Humanoid Physical Intelligence Robot Training Π0.7
United States Latest News, United States Headlines
Similar News:You can also read news stories similar to this one that we have collected from other news sources.
Grading Commanders’ New Uniforms: Washington Blends Past and Future With New ThreadsThe Commanders dropped a brand new uniform set ahead of the 2026 season.
Read more »
Never-before-seen photos reveal new details in murder of New Jersey woman 35 years agoNew details have been revealed in the murder of a woman in New Jersey 35 years ago after a man in South Carolina was arrested last week.
Read more »
New Disclosure Day Footage Reveals Aliens and New MysteriesComic Book Movies, News, & Digital Comic Books
Read more »
Sophia Wilson returns to Colorado with U.S. Women’s National Team with new name, baby girl in towNew name, new mother, new Soph.
Read more »
April 2026 Horoscope: The 5 Signs Starting a New Life Journey During the Aries New MoonOn April 17, the fiery sign of Aries shifts the momentum forward in a big way.
Read more »
New ferry route proposed for World Cup travel between New York and New JerseyA New York City Council member is pushing a proposal to expand ferry service between New York and New Jersey ahead of the upcoming FIFA World Cup
Read more »