GEN-1.5: Embodied Foundation Models are One-Shot Learners is a research post published by Generalist on August 19, 2026, credited to the "Generalist Team" and hosted on the company's own blog (Source: generalistai.com). It announces GEN-1.5, a robot foundation model, and reports that the model acquires new manipulation tasks either from a single demonstration inserted into its context window with no training, or from between one and ten gradient steps on one to five minutes of data. The company characterises these abilities as having emerged from pretraining rather than having been trained for.
The post is a vendor self-report. No independent replication or third-party evaluation accompanied it, and its charts and video demonstrations were published as figures whose underlying values are not stated in the text.
Central claim
The post frames the result by analogy to language models, citing Language Models are Few-Shot Learners (Brown et al., 2020) for the observation that GPT-3 reached roughly 45% average accuracy across a suite of language tasks with one-shot in-context prompting and up to about 65% with roughly 100 examples. Generalist argues that the equivalent capability in robotics — generalising a task from one or a few demonstrations without restriction to particular objects, task types, or sensing modalities — had been pursued since at least the teach-by-guiding of the 1954 Unimate patent and MIT's 1970 Copy Demo, and had "predominantly been considered out of reach."
Against that background the post's claim is that GEN-1.5 is "the first model we know of that has demonstrated the general ability to learn a wide range of dexterous closed-loop physical tasks from just one-shot or few-shot demonstrations." The qualifier is the company's own and is stated as knowledge-limited rather than as a verified priority claim.
Reported results
Four adaptation regimes are reported and are distinguished in the source:
| Regime | Data supplied | Reported success |
|---|---|---|
| One-shot in-context ("physical prompting") | 3–12 seconds of a single demonstration in a 30-second context window; no gradient updates | 59% average (±10% std. dev.) across 10 tasks |
| Few-shot gradient adaptation | 10 gradient steps on 5 minutes of data per task (~50 demonstrations) | 83% average (±9% std. dev.) |
| Few-shot adaptation, general range | 1–10 gradient steps on 1–5 minutes of data (~10–50 demonstrations) | Not stated as a single figure |
| Extreme one-step regime | 1 gradient step, sampling from 1 minute of data | 66.5% on a held-out task |
The ten evaluated tasks include 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. The post states plainly that "the tasks are simple and short-horizon, and the success rates are modest," and that skills learned in context "are currently more brittle than finetuned models." It also reports that in some cases in-context learning of a new task exceeded the performance of one to five gradient steps on the same demonstration data.
Generalist reports that ten adaptation steps change the model's weights on held-out tasks by less than 0.15%, which it reads as evidence that fine-tuning "slightly reconfigures knowledge already present rather than building new representations." It describes the low-step regime as closer to test-time training in an extremely low-data regime than to conventional task-specific training, and notes that the extreme one-step results improve with larger batch sizes and higher learning rates without any adaptation-specific hyperparameter sweep.
Generalisation behaviours reported
Beyond the headline success rates, the post catalogues behaviours it describes as emergent:
- Compositional generalisation. With demonstrations of two independently recorded tasks placed in context — unzipping a pencil pouch and retrieving money — the model chained them into one continuous behaviour, producing intermediate repositioning, regrasping, and error-recovery motions present in neither demonstration. Generalist calls the resulting practice "physical prompt engineering," assembling a compound task from a library of short reusable prompts.
- Zero-shot sim-to-real transfer. A prompt formed entirely from simulated experience drove the real robot, despite pretraining containing no simulation data — neither rendered video nor simulated dynamics. The post distinguishes this from the conventional sense of the term: the model was not trained on the task in either the simulator or the real world.
- Human-to-robot imitation. In some cases a person demonstrating a task with their own hands in view of the robot's cameras was sufficient for the robot to reproduce it.
- Improvised tool use. After fine-tuning on five minutes of human demonstrations of sweeping a block into a bowl with a brush, the model used a banana as a makeshift brush, and when handed a dustpan departed further, using it to lift the block and tip it into the bowl. Generalist states that neither the fine-tuning data nor, to its knowledge, the pretraining data contains a dustpan used this way.
- Other reported behaviours. Removing a Lego brick stuck to its own fingertips with the opposite hand; using two hands to rotate a jar lid where the training data used one; sorting blocks by colour or category when fine-tuned only to place a single block in a bowl; removing a paper covering a bowl and sometimes replacing it; and generalising a lid-twisting skill to unseen cups and bottles.
The post reports that improvisation "strengthens as the number of fine-tuning gradient steps decreases," which it attributes to lightly adapted models staying closer to their pretrained priors.
Model and training as described
GEN-1.5 is described as a large multimodal model that takes video input with 30 seconds of memory, alongside other sensor, language, and proprioceptive inputs, and produces action trajectories at 100 Hz. Physical prompts are sensorimotor examples — sensor data plus action trajectories — recorded either as human data using a pair of handheld grippers or as rollouts from the robot itself.
The company places the model in a lineage: GEN-0 (2025), which it says established predictable scaling laws for the approach, and GEN-1 (2026), which it says could be post-trained to task mastery at 99%-plus success rates and showed initial signs of improvisation. GEN-1.5's pretraining is described as having run continuously for more than eight months across three training phases, on randomly sampled continuous spans of data captured in homes, warehouses, and factories, with no bespoke infrastructure for packing examples into context. A nearest-neighbour language search over the pretraining corpus is described as covering 1,891,392 scenes.
Explanation offered, and its status
Generalist states that it did not train for any of the reported capabilities: no architectural changes to promote in-context learning, no inner or outer meta-learning loop, and no auxiliary objectives encouraging improvisation. It describes the cause as difficult to pinpoint and offers two hypotheses by analogy to language: that the distribution of physical observations and actions may exhibit the "burstiness" and Zipfian structure linked to in-context learning in language models (citing Chan et al., 2022), and that physical work contains naturally repetitive cycles the model may have learned to detect and extend. Both are presented as conjecture, not as findings. The post observes that physical prompts introduce discontinuous jumps in time that the model never encountered during training.
The company's forward-looking statement is that "past a certain threshold of pretraining, the cost of adaptation becomes negligible," and that it does not yet see where the trend asymptotes.
Terminology
"Physical prompting" is Generalist's own term for placing a sensorimotor demonstration in the model's context window; it is used here only in attribution to this source. The same applies to "physical prompt engineering" for the composition of multiple such prompts.
Provenance
Published at generalistai.com on August 19, 2026, on the developer's own domain, and self-cited as "Generalist Team, 'GEN-1.5: Embodied Foundation Models are One-Shot Learners', Generalist AI Blog, Aug 2026." The post cites 22 references, including Brown et al. (2020) on few-shot language learning, Finn et al. (2017) on model-agnostic meta-learning, Bommasani et al. (2021) on foundation models, Akyürek et al. (2024) on test-time training, and Generalist's own earlier GEN-0 and GEN-1 posts. The captured text is complete; figures and embedded video demonstrations are represented by their captions only, so no chart value is recoverable from the capture.
Relationships
- instance-of: AI Robotics — a robotics foundation-model capability report.
- supports: GEN-1.5 — the primary record for the model's reported capabilities.
- depends-on: Scaling Laws — the post frames its result as a consequence of continued pretraining scale.
- instance-of: Emergent capabilities — the capabilities are described as unplanned products of pretraining.
- related: Foundation Models — the paradigm in which the emergence claim is framed.
- related: Physical Intelligence, Figure AI, Helix (Figure AI Vision-Language-Action model), Gemini Robotics 2 — peer embodied foundation-model efforts.