Generalist's GEN-1.5 learns robot tasks from a single 3 to 12 second demonstration, reaching 59% without fine-tuning

Generalist AI said on August 19, 2026 that its newest robot foundation model, GEN-1.5, can learn a new physical task from a single demonstration lasting 3 to 12 seconds, with no gradient updates and no fine-tuning. Across 10 manipulation tasks, the model reached 59% average success when prompted with one example, with a standard deviation of 10 percentage points, the company reported in a research post signed by the Generalist team.

The figures improve quickly once a small amount of training is allowed. With 10 gradient steps on about 5 minutes of data per task, or roughly 50 demonstrations, average success rose to 83%, with a standard deviation of 9 points. In what Generalist calls the extreme one-step regime, a single gradient step on data sampled from one minute of demonstrations produced 66.5% success on a held-out task. The company said 10 steps changed the model's weights on held-out tasks by less than 0.15%, which it reads as the model reconfiguring knowledge it already holds rather than building new representations. Generalist said it did not tune the adaptation procedure or sweep its hyperparameters.

GEN-1.5 is described as a large multimodal model that takes in video with 30 seconds of memory, together with other sensor, language and proprietary inputs, and outputs action trajectories at 100 Hz. Generalist calls its method physical prompting, meaning that sensorimotor examples are placed in the model's context window in the way text examples are placed in a language model's prompt. A demonstration can be recorded by a person using a pair of handheld grippers or taken from the robot's own earlier rollouts, and the company showed a drag and drop interface for choosing which demonstrations go into the context. The tasks shown include putting a marker into a cup, pouring bolts, working zippers, twisting the lid off a glass jar, unzipping a pencil pouch, brushing a cube into a bowl, removing a vacuum pad and taking money out of a wallet.

The company stresses that it did not design the model for this behaviour. According to the post, there were no architectural changes to promote in-context learning, no inner or outer meta-learning loop and no auxiliary objectives. GEN-1.5's pretraining started in parallel with earlier work and has run continuously for more than eight months, and its next-action prediction error on a held-out validation set kept falling across three training phases. "We do not yet see where that curve asymptotes," the team wrote. Generalist also showed a demonstration recorded entirely in simulation being used as a prompt for a real robot, although GEN-1.5's pretraining contained no simulation data, and gave qualitative examples of the robot using either hand after one-handed demonstrations and grasping cups and bottles it had not seen. It published no success rates for these examples.

Generalist frames the result against the history of language models. It cites GPT-3 as reaching roughly 45% average accuracy with one example and up to about 65% with around 100 examples, and positions GEN-1.5's 59% one-shot figure as a comparable early signal. The company says GEN-1.5 is the first model it knows of to show a general ability to learn a wide range of dexterous, closed-loop physical tasks from one or a few demonstrations.

The release is the third in a sequence. GEN-0, published on November 4, 2025, made the company's case for scaling laws in robot learning. GEN-1, released on April 2, 2026, was pretrained on more than 500,000 hours of human physical interaction data collected with wearable devices and no robot data, and reached 99% average success on chosen tasks after about an hour of robot data per task, according to Generalist. On June 4, 2026 the company announced $400 million in new funding led by Radical Ventures, with 8VC, Union Square Ventures, Hanabi Capital and Norwest joining existing investors including NVIDIA, Boldstart Ventures, Spark Capital, Bezos Expeditions and NFDG, taking its total raised above $500 million. It has not disclosed a valuation.

GEN-1.5 landed one day after a rival made a similar pitch. Skild AI published S1 on August 18, 2026, describing it as a foundation model built from the ground up as an in-context learner that executes a task after one video demonstration, seen or unseen and short or long, with no fine-tuning. Skild reported 96% success on in-distribution tasks and 66% on out-of-distribution tasks at 100,000 training hours, and said its long-horizon tasks run up to 10 minutes. The two companies use different tasks, metrics and demonstration formats, so the figures cannot be compared directly. Physical Intelligence, Google DeepMind and NVIDIA are pursuing general robot models through fine-tuning and large pretraining runs rather than a public one-shot benchmark.

For buyers of automation the stakes are practical. Teaching an industrial robot a new job today usually means programming or collecting task data, and GEN-1 had already cut the robot data needed to about an hour per task. If a few seconds of demonstration can replace that hour for simple tasks, the cost of adding a task to a robot cell would fall to the time it takes an operator to show it once. That would matter most for high-mix work in logistics, electronics and light assembly, where tasks change too often to justify long programming projects.

The caveats are spelled out in Generalist's own text. The company calls the tasks simple and short-horizon and the success rates modest, and says skills learned in context are currently more brittle than fine-tuned ones, though they can generalize to some perturbations. The post does not give the number of trials per task, a breakdown of the pretraining data or any independent evaluation. It names no customers, pricing or release date for GEN-1.5. A 59% success rate is far from the reliability factories accept, and the evidence for the model's sim-to-real and cross-hand behaviour is so far qualitative.