Generalist AI Releases GEN-1.5: A Robot Foundation Model That Learns New Tasks From One 3–12 Second Demo

Generalist AI unveiled GEN-1.5, a multimodal foundation model that processes video, proprioceptive, sensor and language streams and emits action commands at 100 Hz. The model retains a 30‑second rolling context window and has been pretrained continuously for over eight months on physical interaction data gathered from homes, warehouses and factories. No architectural modifications were added to encourage in‑context learning; the capability emerged from scale alone.

In a zero‑shot setting, inserting a 3–12 second sensorimotor demonstration into the context window yields an average success rate of 59 % (±10 % std. dev.) across ten manipulation tasks. Performing ten gradient steps on roughly five minutes of data per task improves performance to 83 % (±9 %). A single gradient step on one minute of data achieves 66.5 % on a held‑out task, and the associated weight shift is below 0.15 %, indicating that fine‑tuning reconfigures existing representations rather than learning new ones. After minimal fine‑tuning the model exhibits improvised behavior, such as using a banana as a brush or a dustpan to lift a block, and operates ambidextrously despite single‑handed demonstrations.

GEN-1.5 is not released as open weights or a public API; access is granted only through direct partnership with Generalist AI, which runs the model on its own fleet and data engine. The company describes the release as a research preview, with no pricing or self‑serve options currently available.

Why this matters

The reported results show that a model trained solely on large‑scale physical interaction data can acquire one‑shot task capabilities without any explicit meta‑learning architecture, suggesting that emergent in‑context learning is not exclusive to language domains. The minimal weight change (<0.15 %) after a few gradient steps implies that adaptation in this regime operates by re‑activating latent skills already encoded during pretraining, a form of test‑time training that could drastically reduce the data needed for new robot behaviors. However, the success rates remain modest for complex tasks, and the lack of public weights limits independent verification, so broader applicability remains an open question requiring further study and community access.