Generalist AI model GEN-1.5 teaches robots new tasks with single demo

Generalist AI's GEN-1.5 model teaches robots new tasks using a single demonstration, achieving 59% success on simple tasks and improving to 83% after minimal data.

Generalist AI developed GEN-1.5, a model that teaches robots new tasks through a single demonstration lasting 3–12 seconds. In internal tests across ten simple tasks—like opening jars or withdrawing money—the model achieved an average success rate of 59%. After ten training steps using just five minutes of data, this rate rose to 83%. The model also chains demonstrations, uses simulated environments, and partially replicates human hand movements. Crucially, these capabilities emerged organically during eight months of pretraining without explicit task-specific instruction. Generalist AI claims GEN-1.5 is the first model to demonstrate this across diverse tasks. However, all performance metrics come from internal testing and have not been independently verified. This advancement could reduce labor costs in manufacturing and logistics by enabling cheaper, more adaptable robots. What matters next is whether independent validation confirms these results at scale—especially for more complex tasks beyond the simple demonstrations tested here. The source’s unverified claims mean this remains an experimental step toward practical robotics, not a deployed solution.

Source: The Decoder