Generalist GEN-1.5

Generalist GEN-1.5 is a proprietary robot foundation model announced by Generalist AI on August 19, 2026. Generalist describes it as a large multimodal model that conditions robot control on recent video, other sensor data, language, proprioceptive input, and short sensorimotor demonstrations. The model produces action trajectories at 100 Hz and has a 30-second context window.^[1]

The model's exact name is GEN-1.5; "Generalist GEN-1.5" is a disambiguating article title. The announcement focused on two distinct ways to adapt the model to a new manipulation task. In one-shot in-context use, a single 3- to 12-second demonstration is placed in the context window and the model attempts the task without changing its weights. In a separate fine-tuning mode, Generalist updates the model for 1 to 10 gradient steps using 1 to 5 minutes of task data. The company reported average success rates of 59% for its zero-gradient condition and 83% after 10 gradient steps across ten internal, short-horizon tasks.^[1][2]

GEN-1.5 was introduced through a company research article and demonstration video, not through a downloadable checkpoint, public API, or peer-reviewed paper.^[1][4] Its architecture, data, and evaluation package were not released, and no independent group had reproduced the reported results at announcement. The scores and capability examples are therefore company-reported results rather than public benchmark records.

Release and status

Generalist introduced GEN-1.5 as the next named model in the GEN family after Generalist GEN-1. The company said GEN-1.5's initial pretraining began in parallel with GEN-1 and continued for more than eight months. This does not establish that GEN-1.5 is a fine-tuned version of GEN-1; the public material does not provide a model lineage diagram or checkpoint relationship.^[1][5]

The release calls GEN-1.5 a robot foundation model and a large multimodal model. It does not describe a separate commercial service, product tier, or self-serve interface. A BibTeX entry on the page cites the Generalist blog article itself, but the entry is not a technical paper, software license, or model license.^[1]

Scroll sideways for more →

The official video was uploaded on the same date and runs for 213 seconds. It is a demonstration artifact rather than an independent test.^[4] The release did not name a customer, production deployment, public beta, or GEN-1.5 early-access program. GEN-1 had been offered to selected early-access partners, but that earlier access statement does not by itself apply to GEN-1.5.^[1][5]

Adaptation modes

In-context physical prompting

Generalist uses the term "physical prompting" for inserting a sensorimotor demonstration into the model's context. The demonstration contains observations and action trajectories. It can be recorded by a person operating a pair of handheld grippers or taken from a robot rollout. The remaining portion of the 30-second window holds rolling observations while the robot acts. The company also demonstrated cases in which a person used bare hands in view of the robot's cameras, followed by a robot attempt at the observed task.^[1]

In this mode, the prompt does not trigger fine-tuning or a gradient update. "One-shot" means one demonstration is available in context. It does not mean the robot is evaluated only once, that one model parameter update occurs, or that the task becomes a permanent part of the model. The public release does not show whether a prompted skill persists after the demonstration leaves the context window. It therefore does not establish online learning, continual weight updating, or lifelong retention.^[1]

The phrase "learns in seconds" refers operationally to the short demonstration and immediate conditioned rollout. The task-results chart uses 12 seconds of demonstration context for every listed in-context result, although the broader release says individual prompts can range from 3 to 12 seconds.^[1][2]

Generalist also showed two short demonstrations placed together in the context: opening a pencil pouch and retrieving money. The video depicts a continuous attempt that bridges the two actions with motions not present in either prompt. This is evidence of one selected compositional example. The release does not provide a compositional test set, success rate, or measured limit on the number or duration of prompts that can be combined.^[1]

Few-step fine-tuning

The second mode changes model weights through gradient descent. Generalist reports that GEN-1.5 can adapt with 1 to 10 gradient steps using 1 to 5 minutes of task-specific data, approximately 10 to 50 demonstrations. Its main ten-task comparison uses 10 steps and 5 minutes per task. These are task-specific adapted checkpoints, not the unchanged base model used for the in-context condition.^[1][2]

Generalist says 10 steps changed model weights on held-out tasks by less than 0.15%. The release does not define the percentage's norm or denominator in its prose, so it should not be read as saying that only 0.15% of parameters changed. An accompanying figure uses a two-dimensional multidimensional-scaling projection of pairwise L2 distances between task checkpoints.^[1]

A separate experiment reports 66.5% success after one gradient step using one minute of data on a held-out task. That figure is not the average across the ten tasks in the main chart. The company compared few-step adaptation with test-time training, but did not disclose a deployed loop that continually retrains during ordinary robot operation.^[1]

Architecture and pretraining disclosures

GEN-1.5 accepts video together with unspecified additional sensor, language, and proprioceptive inputs. It outputs 100 Hz action trajectories. The release does not state the parameter count, layer architecture, observation resolution, camera number or models, action coordinate system, action horizon, robot-arm manufacturer, gripper specification, low-level controller, inference hardware, latency, or training compute.^[1]

Generalist says the pretraining data consists of continuous physical activities collected in homes, warehouses, factories, and other settings. Training examples were randomly sampled continuous spans rather than episodes deliberately arranged to teach in-context learning. It also says the GEN-1.5 pretraining mixture contains no rendered simulation video or simulated dynamics. The post does not disclose the total hours, number of contributors, geographic distribution, task distribution, consent process, or held-out-set composition for this model.^[1]

The company plotted declining next-action prediction error on a held-out validation set over more than eight months and three training phases.^[1][3] The chart does not disclose the size of that set or connect a particular validation-error value to task success. Generalist attributes improvements to additional data and compute as well as architectural and algorithmic changes. It separately says it did not add a purpose-built meta-learning loop or an auxiliary objective specifically intended to produce in-context learning or improvisation. Those statements are compatible: the model changed during development, but the company says the observed prompt behavior was not the direct target of a dedicated adaptation objective.^[1]

GEN-1's earlier release described more than 500,000 hours of physical-interaction data.^[5] GEN-1.5's post does not assign that figure, or a replacement total, to its own pretraining corpus. The predecessor's data count therefore cannot be treated as a disclosed GEN-1.5 training total.

Company-reported task evaluation

Generalist evaluated the two adaptation conditions on ten atomic manipulation tasks. The in-context condition used a 12-second prompt and no gradient updates. The fine-tuned condition used 10 steps and 5 minutes of data per task. The exact values below come from the chart data loaded by the official release page.^[2]

Scroll sideways for more →

The task-level values average 58.8% and 82.8%, which the release rounds to 59% and 83%. It labels the aggregates as 59% plus or minus 10% and 83% plus or minus 9% standard deviation.^[1][2] The release does not define whether the deviation is over task rates, model-training runs, or another unit, and it does not report confidence intervals.

Several details needed to reproduce or compare the evaluation are absent. Generalist does not publish the number of attempts behind each percentage, the number of training seeds, initial-state distribution, object-instance splits, camera and environment variation, timeout rules, success-scoring procedure, reset process, intervention policy, or task-selection procedure. It also gives no external policy baseline under the same robot hardware and protocol.^[1][2]

The ten tasks are simple, short-horizon manipulations, a limitation stated by Generalist itself. The results do not measure long-horizon planning, mobile manipulation, work around untrained people, sustained operation, or arbitrary task learning. They also do not establish a production failure rate because the evaluation setting, task distribution, and deployment conditions are different.^[1]

Qualitative generalization demonstrations

The release includes several forms of generalization outside the ten-task table. These examples are qualitative unless otherwise noted.^[1][4]

For simulation-to-real prompting, Generalist placed a demonstration produced in simulation into the context of a real robot. It says the model had not been trained on that task and that its pretraining corpus contained no simulation data. The company reports that prompted behavior transferred to different hands and to changed object positions and sizes for a subset of tasks.^[1] The post does not identify the simulator, enumerate the subset, or provide a success rate, so the example is narrower than a measured general sim-to-real transfer result.

For cross-embodiment prompting, a person performs a task with bare hands where robot cameras can observe it, and the robot then attempts the task. Generalist says this works "in some cases" but gives no evaluation set or mapping method. The evidence does not support a claim that arbitrary human demonstrations can be converted to arbitrary robot embodiments.^[1]

For tool use, Generalist fine-tuned a model for five minutes on demonstrations of brushing a block into a bowl. It then showed the adapted model using a banana as a brush and using a dustpan to lift and pour the block. A language-based nearest-neighbor search over 1,891,392 pretraining scenes found no close example of the dustpan strategy, according to the company.^[1] That search is evidence about the retrieved neighbors, not a complete independent audit proving that no relevant behavior existed anywhere in the proprietary corpus.

Other clips show obstacle removal, use of either hand, two-handed jar manipulation, error recovery, and sorting-like behavior. These selected rollouts illustrate possible strategies but have no reported denominators, control conditions, or failure taxonomy. They do not establish broad emergent capability or safe response to unseen objects.^[1][4]

Relation to prior work

One-shot imitation learning predates GEN-1.5. Duan and colleagues proposed a meta-learning framework in 2017 in which a policy receives one demonstration of a new task and acts on a different instance.^[9] Generalist's novelty claim concerns the reported breadth and scale of physical tasks, not the invention of the one-shot setting.^[1]

Several later systems also conditioned robot policies on demonstrations at inference. The In-Context Robot Transformer used sensorimotor trajectories from human teleoperation to prompt a Franka robot without changing policy parameters, and its authors released code, checkpoints, and data.^[6] Instant Policy reported adaptation from one or two demonstrations using a graph-diffusion method trained with simulated pseudo-demonstrations.^[7]

RoboTTT, published in July 2026, scaled robot-policy context to 8,000 timesteps and reported one-shot imitation from human video.^[8] Its mechanism updates fast weights through gradient descent at inference, which differs from GEN-1.5's zero-gradient in-context mode. These systems use different tasks, robots, data, and evaluation procedures, so their published percentages cannot be directly ranked against Generalist's internal table.

Generalist described GEN-1.5 as the first model it knew to show one-shot and few-shot learning across a broad range of closed-loop physical skills at scale.^[1] No standardized definition of "at scale" or independent priority review accompanied the announcement. The defensible distinction is that Generalist reports the behavior across ten proprietary real-robot tasks without a purpose-built meta-learning objective. A universal first claim is not established by the available evidence.

Reproducibility, safety, and oversight

At announcement, Generalist provided no weights, source code, training data, model card, evaluation scripts, raw logs, or downloadable task package. The public materials therefore did not enable an external lab to run the disclosed model independently, and the internal tasks were not submissions to a public robot benchmark.^[1][2] This contrasts with prior in-context methods such as ICRT and Instant Policy, which published research artifacts.^[6][7]

Real-robot policy evaluation is sensitive to task selection, scene variation, resets, and scoring. RoboArena addressed comparability with more than 600 double-blind pairwise episodes across seven policies and seven academic institutions.^[10] AutoEval automated scene resets and success detection to reduce human evaluation cost.^[11] A 2026 factor-based study used 2,331 real-world evaluations to find failure-prone combinations of object pose, viewpoint, and other conditions.^[12] None of these studies evaluated GEN-1.5; they explain why its internal success rates cannot substitute for a standardized or independently administered test.

The release does not describe a GEN-1.5 safety evaluation, collision or near-miss rates, force limits, object-damage rates, emergency-stop behavior, human-proximity tests, cybersecurity assessment, or fail-safe controller. It also does not state whether a safety operator monitored each rollout or how interventions affected reported success. These omissions do not show that Generalist lacks safety controls. They mean the public evidence cannot support claims of safety certification, unsupervised operation around people, household readiness, or production deployment.^[1]

References

1. ^Generalist Team. GEN-1.5: Embodied Foundation Models are One-Shot Learners. Generalist AI, August 19, 2026.
2. ^Generalist AI. GEN-1.5 task results chart source. August 19, 2026.
3. ^Generalist AI. GEN-1.5 pretraining validation plot source. August 19, 2026.
4. ^Generalist. Introducing GEN-1.5, a one-shot learner. YouTube, August 19, 2026.
5. ^Generalist Team. GEN-1: Scaling Embodied Foundation Models to Mastery. Generalist AI, April 2, 2026.
6. ^Fu, Letian; Huang, Huang; Datta, Gaurav; et al. In-Context Imitation Learning via Next-Token Prediction. arXiv:2408.15980, 2024.
7. ^Vosylius, Vitalis; Johns, Edward. Instant Policy: In-Context Imitation Learning via Graph Diffusion. arXiv:2411.12633, revised 2025.
8. ^Jiang, Yunfan; Chebotar, Yevgen; Zheng, Ruijie; et al. RoboTTT: Context Scaling for Robot Policies. arXiv:2607.15275, July 2026.
9. ^Duan, Yan; Andrychowicz, Marcin; Stadie, Bradly C.; et al. One-Shot Imitation Learning. arXiv:1703.07326, 2017.
10. ^Atreya, Pranav; Pertsch, Karl; Lee, Tony; et al. RoboArena: Distributed Real-World Evaluation of Generalist Robot Policies. Proceedings of the 9th Conference on Robot Learning, PMLR 305:336-364, 2025.
11. ^Zhou, Zhiyuan; Atreya, Pranav; Tan, You Liang; et al. AutoEval: Autonomous Evaluation of Generalist Robot Manipulation Policies in the Real World. arXiv:2503.24278, 2025.
12. ^Liao, Andrew; Cui, Hanchen; Desingh, Karthik; Deshwal, Aryan. Active Real-World Factor-Based Evaluation for Generalist Robot Policies. arXiv:2607.14439, July 2026.