San Francisco-based startup Physical Intelligence has unveiled new research indicating that its latest AI model can enable robots to perform tasks they were never explicitly trained to do. The development marks a potential step toward creating general-purpose robotic systems capable of adapting to unfamiliar environments.

The model, known as π0.7, demonstrates what researchers describe as “compositional generalization,” allowing machines to combine previously learned skills to solve new problems. This represents a shift from traditional robotics training, which typically relies on task-specific programming.

In one experiment, the system successfully interacted with an unfamiliar appliance by synthesizing limited prior data with broader knowledge. Researchers also found that simple human guidance, delivered through natural language instructions, significantly improved the robot’s performance.

Despite the promising results, experts caution that the technology remains in its early stages. While the model has shown the ability to perform complex activities such as assembling objects and preparing simple tasks, it still requires step-by-step instruction and lacks full autonomy in multi-stage operations.

Source: TechCrunch