Generalist is a robotics company developing embodied foundation models under the GEN name. Its stated approach is to treat physical manipulation skills as something a pretrained model can acquire in context, from a demonstration supplied at inference time, rather than through task-specific fine-tuning.
GEN-1.5
Generalist published GEN-1.5 on August 19, 2026, describing it as an embodied foundation model that learns new physical tasks from a single demonstration of three to twelve seconds placed in a 30-second context window, with no gradient updates or fine-tuning — an approach the company calls "physical prompting" (GEN-1.5: Embodied Foundation Models are One-Shot Learners).
Across ten manipulation tasks — including twisting a lid off a glass jar, unzipping a pencil pouch and retrieving money from a purse — the company reports the pretrained model averaging 59% success one-shot, with a standard deviation of 10 percentage points, rising to 83% (±9) after ten gradient steps on roughly five minutes of data per task, about 50 demonstrations. Generalist states that the model has been pretraining continuously for over eight months, that it was not trained explicitly for in-context learning, and that ten adaptation steps change the model's weights on held-out tasks by less than 0.15% (GEN-1.5: Embodied Foundation Models are One-Shot Learners).
The company qualifies the result in two ways: the evaluated tasks are simple and short-horizon, and skills acquired in context are more brittle than fine-tuned ones (GEN-1.5: Embodied Foundation Models are One-Shot Learners).
A competing claim followed within a week. Skild AI released S1 on August 18, 2026, also an in-context manipulation model, and its post names GEN-1.5 as a concurrent approach that it argues is "constrained to short horizons, in-distribution tasks, or both" (Introducing S1: In-Context Learning for Robotics). The two sets of figures are not directly comparable: Generalist reports average task success over ten primitives, Skild an average per-step success rate on longer unseen tasks under a protocol that uses human intervention to recover from failures. Both are vendor self-reports on internal benchmarks.
GEN-1.5 is the third publicly described model in the GEN series. The company describes GEN-0, announced in 2025, as the point at which predictable scaling laws for embodied pretraining became visible, and GEN-1, announced in 2026, as post-trainable to task mastery at success rates above 99%. GEN-1.5's pretraining began in parallel with the GEN-1 announcement and had run continuously for more than eight months at publication (GEN-1.5: Embodied Foundation Models are One-Shot Learners).
All figures are the company's own and were not independently verified at publication; no benchmark comparison to peer robotics foundation models has been published alongside them.
Relationships
- instance-of: AI Robotics — one of the robotics foundation-model efforts.
- related: GEN-1.5 — the company's current embodied foundation model.
- supports: GEN-1.5: Embodied Foundation Models are One-Shot Learners — the company's own account of the model.
- related: Physical Intelligence, Figure AI, Helix (Figure AI Vision-Language-Action model) — peer robotics foundation-model developers and models.
- related: Embodied AI vs AGI (China's Race Redirect) — the in-context-adaptation claim bears on arguments about what embodiment adds to generality.