GEN-1.5 is a robot foundation model developed by Generalist and announced on August 19, 2026 in the research post "GEN-1.5: Embodied Foundation Models are One-Shot Learners". It is the third publicly described model in the company's GEN series, following GEN-0 (2025) and GEN-1 (2026). Its distinguishing claim is that it acquires new manipulation tasks from a demonstration supplied at inference time and held in its context window, without gradient updates — a mode the company terms "physical prompting."
All published figures are Generalist's own. As of August 25, 2026 no independent evaluation, benchmark-leaderboard entry, or third-party replication had been reported, and the model has no published system card or safety case.
Architecture and inputs
Generalist describes GEN-1.5 as a large multimodal model that takes video input with 30 seconds of memory, together with other sensor, language, and proprioceptive inputs, and emits action trajectories at 100 Hz (GEN-1.5: Embodied Foundation Models are One-Shot Learners). Parameter count, model family, and hardware are not disclosed.
A "physical prompt" is a sensorimotor example — sensor data paired with an action trajectory — inserted into the context window, with the remainder of the window holding rolling observations. Prompts are recorded either as human data captured with a pair of handheld grippers, or as rollouts from the robot itself.
Training
Pretraining ran continuously for more than eight months across three training phases and began in parallel with the release of GEN-1. The corpus is drawn from the company's data engine — physical activity captured in homes, warehouses, and factories — sampled as random continuous spans, with no bespoke infrastructure for packing discrete examples into context. The company reports a nearest-neighbour language search covering 1,891,392 scenes in that corpus. Pretraining contains no simulation data, neither rendered video nor simulated dynamics (GEN-1.5: Embodied Foundation Models are One-Shot Learners).
Generalist states that none of the model's in-context or few-shot abilities were trained for: no architectural change to promote In-Context Learning, no inner or outer meta-learning loop, and no auxiliary objective encouraging improvisation. Skild AI's S1, released the day before GEN-1.5, takes the opposite position on training, describing pre-training conditioned only on in-context demonstrations as an outer meta-learning loop that teaches the policy how to learn from context (Introducing S1: In-Context Learning for Robotics).
Capabilities and reported results
Four adaptation regimes are distinguished in the announcement and are easily conflated:
| Regime | Data supplied | Reported success |
|---|---|---|
| One-shot in-context | 3–12 s of one demonstration, no gradient updates | 59% average (±10% std. dev.), 10 tasks |
| Few-shot, 10 steps | 10 gradient steps on 5 min/task (~50 demonstrations) | 83% average (±9% std. dev.) |
| Few-shot, general range | 1–10 gradient steps on 1–5 min (~10–50 demonstrations) | No single figure stated |
| Extreme one-step | 1 gradient step on 1 min of data | 66.5% on a held-out task |
The evaluated tasks are manipulation primitives: twisting a lid off a glass jar, unzipping a pencil pouch, brushing a cube into a bowl, removing a vacuum pad, and retrieving money from a wallet. Generalist states that the tasks are "simple and short-horizon" and the success rates "modest," and that in-context skills are more brittle than fine-tuned ones. In some cases in-context learning exceeded one to five gradient steps on the same data (GEN-1.5: Embodied Foundation Models are One-Shot Learners).
The company reports that ten adaptation steps move the model's weights on held-out tasks by less than 0.15%, and frames the low-step regime as closer to test-time training in a very low-data regime than to conventional fine-tuning.
Generalisation behaviours
Generalist reports several behaviours it describes as emergent rather than trained:
- Composition of prompts. Two independently recorded demonstrations placed in context together were chained into one continuous behaviour, with bridging motions — repositioning, regrasping, error recovery — present in neither prompt.
- Sim-to-real prompting. A prompt formed entirely in simulation drove the real robot, despite the absence of simulation data in pretraining. The company distinguishes this from conventional zero-shot sim-to-real transfer, since the model was trained on the task in neither setting.
- Human-to-robot imitation. A person demonstrating with their own hands in view of the robot's cameras was in some cases sufficient for immediate reproduction.
- Improvised tool use. Fine-tuned on five minutes of brush-sweeping demonstrations, the model used a banana as a substitute brush, and used a dustpan to lift and tip the block rather than sweep it. Generalist states that neither the fine-tuning data nor, to its knowledge, the pretraining data contains a dustpan used this way.
- Further reported behaviours. Removing a Lego brick stuck to its fingertips with the opposite hand; bimanual lid rotation where training used one hand; sorting blocks by colour or category when fine-tuned only to place one block in a bowl; removing and sometimes replacing a paper covering the target bowl; and transferring a lid-twisting skill to unseen cups and bottles.
Improvisation is reported to strengthen as the number of fine-tuning steps decreases, which the company attributes to lightly adapted models remaining nearer their pretrained priors.
Explanation offered for the capability
Generalist states that it cannot pinpoint why in-context learning emerges. It offers two hypotheses by analogy to language models — that physical observations and actions may exhibit the "burstiness" and Zipfian structure linked to in-context learning in language (citing Chan et al., 2022), and that physical work contains repetitive cycles the model may have learned to detect and extend. Both are presented as conjecture. The company notes that physical prompts introduce discontinuous jumps in time that the model never saw in training (GEN-1.5: Embodied Foundation Models are One-Shot Learners).
Lineage
GEN-0, announced in 2025, is described by the company as the point at which predictable scaling laws for embodied pretraining became visible. GEN-1, announced in 2026, is described as post-trainable to task mastery at success rates above 99%, with initial signs of improvisation. Neither has a dedicated page. GEN-1.5's pretraining began in parallel with GEN-1's announcement (GEN-1.5: Embodied Foundation Models are One-Shot Learners).
Availability
No public API, weights, pricing, or general availability had been announced as of August 25, 2026. The announcement is a research post rather than a product launch, and demonstrations are shown on the company's own robot hardware.
Comparison with peer models
GEN-1.5 sits among embodied foundation models pursuing generality across tasks rather than mastery of one: Helix from Figure AI, which pairs a vision-language-action model with the developer's own humanoid; Gemini Robotics 2 from Google DeepMind; and the embodiment-flexible models of Physical Intelligence. The published claims differ in kind rather than degree — GEN-1.5's are about adaptation cost from a demonstration, where the others' are largely about task breadth and instruction following — and no common benchmark places them on one scale.
Relationships
- developer: Generalist
- instance-of: AI Robotics
- instance-of: Emergent capabilities — the developer states it trained for none of the reported abilities.
- depends-on: Foundation Models, Scaling Laws — the company frames the result as a product of continued pretraining scale.
- related: S1 (Skild AI) — the concurrent in-context manipulation claim, released the day before.
- related: Helix (Figure AI Vision-Language-Action model), Gemini Robotics 2, Physical Intelligence, Figure AI — peer embodied foundation-model efforts.
- related: Embodied AI vs AGI (China's Race Redirect) — the adaptation claim bears on what embodiment contributes to generality.
Sources
- GEN-1.5: Embodied Foundation Models are One-Shot Learners — Generalist, "GEN-1.5: Embodied Foundation Models are One-Shot Learners," August 19, 2026.