A three-to-twelve-second demonstration video, paired with motion data recorded alongside it, gets stuffed into the model's context window, and the robot then goes and performs a task it has never practiced — no gradient updates, no fine-tuning, not a single weight moves. Generalist AI calls this approach "physical prompting," unveiled alongside its new model, GEN-1.5.
GEN-1.5 is a large multimodal model that takes four kinds of input — video, sensor readings, language instructions, and proprioception — retains a 30-second memory window, and outputs action trajectories at 100 Hz. The team says it has been training continuously on its data engine for more than eight months.
A range bounded by three numbers
The company tested the approach on 10 manipulation tasks — opening a glass jar, unzipping a pencil case, pulling cash out of a wallet, and similar chores. With a single demonstration prompted in context, the average success rate was 59% (±10% standard deviation). Switch to five minutes of data collection per task — roughly 50 demonstrations — followed by 10 steps of gradient updates, and the score climbs to 83% (±9%). On held-out tasks with one step of adaptation, it lands at 66.5%.
That 24-point gap costs five minutes of data collection plus a round of training queue time. For a production line that swaps between dozens of parts, the math isn't hard to work out: the fine-tuning route means running through collection and training every time the SKU changes, while in-context prompting takes seconds. Still, 59% is a long way from unattended operation — for now it looks more like a way to skip the reprogramming step in human-robot collaboration.
Set this against the approach favored by Chinese peers. Companies such as Alibaba, AgiBot, and Unitree have spent the past year mainly scaling up data and refining hardware embodiments, pushing single-task success rates higher through massive teleoperation datasets — but still have to redo collection and training for every new task. GEN-1.5 chose a different trade: giving up peak single-task scores in exchange for the freedom to switch tasks without touching the weights. Rough math suggests that for a factory bringing three to five new workstations online each week, the two approaches could differ by an order of magnitude in deployment labor.
Even the team was caught off guard
"Inserting a single demonstration in the context buffer... yields any measurable competence at all was unexpected."
That line captures the weight of the finding. In-context learning in language models also "emerged" back in the GPT-3 era, without anyone training for it specifically; what Generalist is now reporting looks like the same kind of emergence showing up on the sensor-and-joint-angle side. The company was founded in 2024 by former DeepMind and Boston Dynamics researchers, and has previously raised $400 million in funding.
The limits the company drew itself
The announcement doesn't dodge the shortcomings: the test tasks are all short-horizon and simple manipulations, so the success rate isn't dazzling; skills picked up in context are far more brittle than those from a fine-tuned model; and physical prompting introduces timeline jumps the model never saw during training, leaving discontinuities between the demo clip and actual execution.
Taken together, these limitations point to the same conclusion: learning a new skill from a single demonstration has only just poked its head above water. Whether it can scale to longer-horizon tasks, or hold up in cluttered environments, depends on how far the next few versions can push that 59% figure.
Sources: Generalist AI's official GEN-1.5 announcement blog post, IoT Tech News, CocoLoop; the 59% single-demo figure, 83% post-fine-tuning figure, 66.5% held-out-task figure, 30-second memory window, and 100 Hz action rate all follow the official announcement's figures.