Generalist AI unveils gen-1.5 robot model that learns tasks from a single brief demo

Generalist AI released GEN-1.5, a robot foundation model that can learn a new physical task from a single demonstration lasting 3 to 12 seconds. Sensor and motion data are fed into a 30-second context window, and the robot executes the task immediately. The approach requires no gradient updates, no fine-tuning, and no task-specific code. The company calls the mechanism “physical prompting” and stresses that it involved no architectural changes, no meta-learning loop, and no auxiliary objectives. The capability emerged after more than eight months of continuous pre-training on physical interaction data collected in homes, warehouses and factories.

GEN-1.5 is a large multimodal model that ingests video, sensor streams, language and proprioception, retains a 30-second memory, and outputs action trajectories at 100 Hz. Sensorimotor streams together with the predicted action trajectory are pulled into the context window via a drag-and-project interface; the remainder of the window contains rolling observations. The model produces the task instantly, with zero gradient steps. Generalist AI notes that the tasks are simple short-horizon, but claims this is the first model the team knows that exhibits one-shot learning of physical skills at scale.

Across ten diverse manipulation tasks, a single-shot physical prompt achieved an average success rate of 59 % (±10 %) directly from the pretrained model. Ten gradient steps on five minutes of task-specific data—about 50 demonstrations—raised performance to 83 % (±9 %). In an extreme case, a single gradient step on one minute of data reached 66.5 % on a held-out task without dedicated hyper-parameter search. Typically, adapting a robot policy requires tens of thousand of gradient steps; here ten steps moved model weights on held-out tasks by less than 0.15 %, suggesting that fine-tuning merely re-organizes existing knowledge rather than building new representations. Generalist AI frames this as “test-time training” in an extreme low-data regime.

Three transfer results merit attention. First, compositional generalization: two separately recorded prompts, when concatenated in the context, yield a continuous behavior that includes bridging motions, re-positioning, re-grasping and error recovery—none of which appear in any single demonstration. Second, zero-shot sim-to-real: a demonstration recorded entirely in simulation serves as a prompt for a real robot, even though pre-training never saw simulation data, random video footage or synthetic dynamics. For some tasks, physical demonstrations are no longer required. Third, human-to-robot imitation: in certain cases a human demonstrates a motion in view of the robot’s cameras, and the model reproduces the action with the robot’s arm.

Even after a brief fine-tuning, the model shows generalization. After five minutes of training on pushing a block into a bowl with a brush, the model used a banana as an improvised brush and a dustpan to… (text truncated). Generalist AI emphasizes that this is a research version: there are no public weights, no API, no pricing page and no self-serve product. GEN-1.5 runs on the company’s own hardware and data engine, and access today is available only through direct partnership.