Apple researchers show generalist AI agent outperforming specialist robot models

A new paper from Apple researchers introduces AMP1, a system that lets a single generalist AI model control very different types of robots, including single-arm, two-armed, and mobile manipulators, without needing demonstrations for each one. In testing on a benchmark called RoboLab-120, the generalist approach succeeded on its first try 85.6 percent of the time, far ahead of leading specialized robot-control models that scored around 49 and 43 percent. In real-world trials spanning twelve different tasks and three physical robots, the system succeeded 104 out of 108 times. The researchers also found that adding a memory component, which lets the system recall past attempts, cut failure rates by more than four times on a separate benchmark. The results suggest that general-purpose AI models, rather than robot-specific trained systems, may become the more practical foundation for controlling varied physical robots in real environments.

Why it matters

Suggests general AI models could replace robot-specific training, simplifying how companies deploy robots across different hardware.

Source: arXiv