GEN-1.5 Robot Model Learns Tasks From 3–12 Second Demos

A robot that learns from a single demonstration

Robotics has long promised machines you could teach by showing them a task instead of programming it. In August, Generalist AI got closer to that goal than most of its rivals. The San Mateo startup released GEN-1.5, a robot foundation model that tries a new manipulation task after seeing one demonstration of 3 to 12 seconds. It does this without any gradient updates or fine-tuning.19 The company calls the technique "physical prompting." The name is a deliberate nod to the in-context learning that turned large language models into general-purpose tools.4

The headline numbers are respectable but not dramatic, and that is the right way to read them. Generalist tested the model on 10 physical tasks. With one demonstration, it succeeded 59% of the time on average. When Generalist added five minutes of task-specific data and ran 10 gradient steps, the average rose to 83%.12 The tasks were deliberately short and simple: twisting a lid off a glass jar, pulling money from a purse, stacking cups, sweeping trash, opening a book, unzipping a pencil pouch and removing a vacuum pad.1 Generalist itself calls the tasks simple and short-horizon. It also says this is the first model it knows of in which one-shot and few-shot learning of physical skills has emerged at scale.9

The important part is how quickly the robot picks up a task, not how often it succeeds. A three-in-five success rate would not survive on a factory line. But a robot that gets to roughly 60% from a few seconds of data, and passes 80% with five minutes more, changes how the cost of automation works.

How physical prompting works

GEN-1.5 is a large multimodal model. It takes in video, sensor readings, language and proprioceptive data (the robot's sense of its own body position), keeps 30 seconds of context, and outputs action trajectories at 100 Hz.110 A demonstration goes into that context window as a sensorimotor example, meaning sensor streams plus the action trajectory. The rest of the window holds live observations. Demonstrations can come from a person using a pair of handheld grippers or from the robot's own earlier runs.9 Operators choose which examples to insert through a drag-and-drop interface.3

The coverage differs slightly on what the demonstration actually is. Some outlets described the model as learning from short "demonstration videos."6 Generalist's own description is more specific: the prompt includes action data, not just pixels.9 The difference matters. A model that needs a recorded trajectory from instrumented grippers is a different product from one that learns by watching a phone video. In some cases, though, Generalist says a person demonstrating with bare hands in view of the robot's cameras was enough for the robot to copy the action with its own hands.13

Generalist's strongest scientific claim is that none of this was designed in. The company says it made no architectural changes, used no meta-learning loop and added no auxiliary objectives to encourage in-context learning. In its account, the ability emerged from more than eight months of continuous pretraining.310 It has a tentative explanation. Physical interaction data may share the "bursty," uneven statistical patterns that research has linked to in-context learning in language models. Or the repetitive cycles of physical work may have taught the model to spot patterns and continue them.4

One detail supports the idea that the model is drawing on what it already knows. Ten gradient steps change the weights by less than 0.15%, which suggests fine-tuning is rearranging existing knowledge rather than building new representations.3 One analysis noted that adapting robot policies has usually taken tens of thousands of gradient steps.3

Beyond imitation

Several results suggest the model is doing more than replaying a recorded motion. When given separate demonstrations of unzipping a pouch and retrieving money from it, GEN-1.5 joined them into one continuous sequence. It added repositioning, regrasping and recovery motions that appeared in neither demonstration.39 A demonstration recorded entirely in simulation also worked as a prompt for the real robot, even though the pretraining data contained no simulation at all.310

The tool-use examples got the most attention. After a short fine-tune on sweeping a block into a bowl with a brush, the robot used a banana as a substitute brush. Given a dustpan instead, it switched to scooping the block up and tipping it into the bowl.12 It also cleared away a sheet of paper covering the bowl.3

The outside reaction was enthusiastic. Ted Xiao, who previously worked on Gemini Robotics at Google DeepMind, called one-shot task learning a long-sought "holy grail" and said GEN-1.5 was the best result yet in that direction. He compared the feeling to first using GPT-3.6

The caveats are in the fine print

On the limits, the reporting is consistent, and readers should take those limits seriously. Every key number comes from Generalist's own evaluations. No paper gives the model's parameter count or full training recipe.6 Performance on deformable objects, long tasks and completely unfamiliar environments has not been tested.6 The weights have not been published, and there has been no independent check of safety.5 The company also admits that skills learned through prompting are more brittle than skills learned through full fine-tuning.45

GEN-1.5 is also not a product yet. There are no public weights, no API and no pricing. Access goes through direct partnerships, and Generalist runs the model on its own fleet.3 One profile put the practical question well. A one-shot demo works best when tasks are short and the occasional failure is acceptable. A production line needs long sequences, safety limits and the same quality across thousands of cycles.11

The "emergence" claim also deserves caution. In language models, whether capabilities truly emerge at scale has been debated for years. GEN-1.5's results will need independent replication on harder tasks before that claim is settled.4

Data is the real moat

For the robot foundation model race, the bigger story is data. There is no open corpus of robot manipulation data comparable to the web text that trained language models. Generalist responded by building its own handheld gripper devices and sending them to data contributors around the world.4 Its GEN-0 announcement in November 2025 cited more than 270,000 hours of real interaction data. By the time of GEN-1.5, coverage put the pretraining set above 500,000 hours, collected in homes, warehouses and factories.114

If Generalist is right that in-context learning appeared once its dataset grew large enough, then the company that owns the largest and most varied physical dataset holds a compounding advantage. That reading fits the company's fundraising. Generalist raised $400 million in June at a $2 billion valuation, with Nvidia, Bezos Expeditions and Union Square Ventures among its backers.56 The founders include alumni of Google DeepMind and Boston Dynamics.6

A crowded race toward the robot "prompt"

Generalist is not alone. Weeks later, Skild AI launched S1, which it calls a robot foundation model that learns previously unseen, long-horizon tasks from a single video demonstration, also without weight updates.1315 S1 handled tasks lasting up to 10 minutes, including potting a plant, making pancakes and brewing pour-over coffee. On new multistep tasks it succeeded about 66% of the time at each step, compared with 9% for a comparable system.1516 Skild also reports commercial results: a $100 million annual revenue run rate 10 months after its first deployment, plus a Foxconn and Nvidia project assembling Blackwell systems with dual-arm robots.1518 Like Generalist's numbers, Skild's figures come from the company and have not been independently validated.18

The two releases point in the same direction from different starting points. Generalist reports cleaner one-shot learning on short, atomic tasks. Skild emphasizes longer tasks and commercial use, though a 66% success rate per step multiplies into a modest chance of completing a 20-step job without error.15 The academic background goes back further. Methods such as RoboCLIP used pretrained video-language models to learn from a single demonstration, but noted that long, multi-stage tasks remained difficult.20

What it means for humanoids

GEN-1.5 runs on robot arms, not humanoids. Its importance for the humanoid sector is mostly indirect, but still large. Humanoid makers have been selling general-purpose bodies on the expectation that software will make them useful across many jobs. Generalist's results suggest the most valuable part may be a model that works across many different bodies. The company reported about 9,000 end-effector variations in its GEN-1 training data.11 That points to a hardware-agnostic approach rather than a bet on one form factor.

The conclusion here is that GEN-1.5 is a real signal that scaling works in robotics, not a deployment breakthrough. It shows that pretraining on large amounts of physical data can produce new qualitative abilities, not just small gains. The core claims still need replication, harder benchmarks and longer tasks before anyone can call this robotics' GPT-3 moment. The company that closes the gap between 59% and dependable performance first will have the real business.