Generalist AI introduces GEN-1.5, a robot model that master new tasks from brief demonstrations

Generalist AI has unveiled GEN-1.5, a robot foundation model capable of learning new physical tasks from a single demonstration. The model processes 3–12 seconds of sensorimotor data within a 30-second context window, enabling it to perform the task without gradient updates, fine-tuning, or task-specific programming. Across 10 diverse manipulation tasks, the model achieved an average success rate of 59% (±10% std. dev.) using only the pretrained version, which rose to 83% (±9%) after ten gradient steps on five minutes of data per task. The company refers to this method as physical prompting, emphasizing that it emerged naturally from over eight months of continuous pretraining on physical interaction data, without architectural changes, meta-learning loops, or auxiliary objectives. The model integrates video, sensor, language, and proprioceptive inputs, and emits 100 Hz action trajectories. It runs exclusively on Generalist AI’s internal fleet and data engine, with no public weights, API, or pricing available. The company highlights that the model’s ability to generalize, even in novel scenarios such as using a banana as a makeshift brush or a dustpan to lift a block, suggests a significant leap in one-shot learning for physical tasks. Generalist AI notes that adapting robot policies typically requires tens of thousands of gradient steps, but GEN-1.5 achieves similar results with just ten steps, moving model weights by less than 0.15% on held-out tasks. This approach is framed as test-time training in a low-data regime. The model’s performance also improves with light fine-tuning, demonstrating its adaptability. Generalist AI has shared its research post and announcement thread, and invites collaboration for promoting GitHub repositories, Hugging Face pages, product releases, or webinars. The post originally appeared on MarkTechPost.