Generalist AI Releases GEN-1.5: A Robot Foundation Model That Learns New Tasks From One 3–12 Second Demo
Generalist AI has launched GEN-1.5, a robotic basis mannequin that learns a brand new bodily activity from a single demonstration. Drop 3–12 seconds of sensorimotor knowledge into its 30-second context window, and the robotic performs the duty. No gradient updates, no fine-tuning, no task-specific programming. Across 10 various manipulation duties, this one-shot in-context prompting averaged 59% success (±10% std. dev.) straight from the pretrained mannequin. Ten gradient steps on 5 minutes of information per activity raised that to 83% (±9%). Generalist calls the mechanism bodily prompting, and says it was by no means skilled for: no architectural modifications, no meta-learning loop, no auxiliary goals. It emerged from over eight months of steady pretraining on bodily interplay knowledge. The duties are easy and short-horizon, and the corporate says so plainly. But that is the primary mannequin its group is aware of of the place one-shot studying of bodily abilities has emerged at scale.
Is it deployable?
Not but — it is a analysis launch. There are not any public weights, no API, no pricing web page and no self-serve product. Generalist AI runs GEN-1.5 by itself fleet and knowledge engine. Anyone who needs it at this time goes by way of a direct partnership.
What is GEN-1.5?
GEN-1.5 is a big multimodal mannequin that takes video, sensor, language and proprioceptive inputs, holds 30 seconds of reminiscence, and emits 100 Hz motion trajectories. It has been pretraining repeatedly for over eight months on bodily interplay knowledge captured in houses, warehouses and factories.
The predominant mechanism is bodily prompting. A sensorimotor instance — sensor streams plus the motion trajectory — is inserted into the 30-second context window by way of a drag-and-drop interface. The the rest of the window holds rolling observations. The mannequin then performs the duty instantly, with zero gradient steps and no fine-tuning.
Crucially, none of this was designed in. Generalist states there have been no architectural modifications to advertise in-context studying, no meta-learning loop, and no auxiliary goals encouraging improvisation. The functionality emerged from pretraining scale, the identical approach one-shot prompting emerged in GPT-3.
The numbers
Across 10 various duties, one-shot in-context prompting averaged 59% success (±10% std. dev.) from the pretrained mannequin, with no coaching in any respect. Ten gradient steps on 5 minutes of information per activity — roughly 50 demonstrations — raised that to 83% (±9%). In the acute case, one gradient step on one minute of information reached 66.5% on a held-out activity, with no adaptation-specific hyperparameter sweep.
The compute story is the attention-grabbing half. Adapting robotic insurance policies has sometimes taken tens of hundreds of gradient steps. Ten steps right here transfer the mannequin weights on held-out duties by lower than 0.15%, which suggests fine-tuning is reconfiguring data the mannequin already has quite than constructing new representations. Generalist frames it as test-time training in a particularly low-data regime.
Three switch outcomes value understanding
- Compositional generalization: Two independently recorded prompts positioned in context get chained into one steady behaviour. The mannequin produces the bridging motions — repositioning, regrasping, error restoration — that seem in neither demonstration.
- Zero-shot sim-to-real: A demonstration recorded fully in simulation works as a immediate for the actual robotic, regardless of pretraining containing no simulation knowledge — neither rendered video nor simulated dynamics. For some duties, demonstrations not should be collected bodily.
- Human-to-robot imitation: In some circumstances an individual demonstrates with their very own fingers, in view of the robotic’s cameras, and the mannequin reproduces it with the robotic’s fingers.
Generalization additionally reveals up after gentle fine-tuning. Trained on 5 minutes of brushing a block right into a bowl, the mannequin used a banana as a makeshift brush, and used a dustpan to raise and dump the block as a substitute — a distinct contact sequence fully. It additionally eliminated a sheet of paper masking the bowl, and labored ambidextrously when demonstrations used one hand.