Generalist Unveils GEN-1 General-Purpose AI Model for Physical Systems

To develop GEN-1, Generalist enhanced training stability, created custom kernels, devised novel paged attention methods for real-time inference, refined post-training techniques, and improved controls for smoother, more precise operation. | Source: Generalist AI

Generalist AI Inc. unveiled its GEN-1 general-purpose AI model for robotics yesterday. The company states the system boosts average task success rates to 99%, up from 64% with prior models. It also completes tasks approximately three times faster than current methods, achieving these results with just one hour of robot data per task, according to Generalist.

Founded in 2024, the company develops embodied foundation models for versatile robots. Based in San Mateo, California, Generalist claims GEN-1 "unlocks commercial viability across a wide spectrum of applications." This release follows just five months after its GEN-0 model, which the company said confirmed the existence of scaling laws in robotics.

While optimistic about the AI model's advancement, Generalist acknowledged that GEN-1 cannot solve every task. The startup added that some real-world applications would require success rates exceeding 99% to be practically useful.

Editor’s note: Sessions on embodied and physical AI development will be featured at the 2026 Robotics Summit & Expo on May 27 and 28 in Boston. Registration is now open.

GEN-1 trains on real-world data, scaling up from GEN-0

GEN-1 builds upon GEN-0's foundation through further scaling and algorithmic improvements to begin mastering basic tasks, Generalist AI explained. The model was trained from the ground up on the company's dataset comprising half a million hours of real-world data.

With GEN-0, Generalist demonstrated that robot learning can be scaled in a generalized manner, similar to the predictable progress seen in language models. The company noted that every zero-shot task it monitored showed simultaneous improvement. However, it admitted the model's performance "was not sufficient for commercial deployment."

GEN-1 is the result of increased data and computational scaling, accelerated by algorithmic breakthroughs, stated Generalist. The company reports that some tasks are now reaching the performance threshold required for economically viable real-world use.

The company pointed out that previous general robotics models achieving over 90% success relied on massive, expensive, and difficult-to-scale teleoperation datasets. In contrast, the base foundation model for both GEN-0 and GEN-1 is trained without any robot-specific data.

Instead, the model utilizes data from low-cost wearable devices worn by humans performing millions of activities. Generalist says it has proven this pretraining approach can lead to high mastery levels without needing large teleoperation or simulation datasets.

Generalist leverages advancements across multiple technologies

According to Generalist AI, GEN-1 incorporates pretraining innovations that enhance computational efficiency. Advances in post-training techniques, learning from experience (reinforcement learning), multimodal human guidance, and new inference-time methods also contributed to higher performance on any given task.

Beyond these improvements, the company stated GEN-1 represents a significant increase in computational scale compared to its predecessor. "It demonstrated the ability to quickly learn new tasks, adapt to new environments, and exhibit moments of physical common sense," Generalist noted.

The company claims GEN-1 is a data-efficient learner. In certain tests, the model achieved performance comparable to GEN-0 using ten times less task-specific data and fewer fine-tuning steps.

Since the pretraining dataset contains no robot data, when GEN-1 adapts to a new task, it is simultaneously learning both the specific robot embodiment and the task itself for the first time, Generalist explained.

GEN-1 enhances reliability and improvisational intelligence

"Embodied foundation models must be reliable, fast, and capable of recovering from unexpected situations," said Generalist. Regarding reliability, the company stated GEN-1 can perform several tasks at high reliability levels over extended periods without human intervention.

The company demonstrated GEN-1 across six tasks: kitting auto parts for over an hour, folding T-shirts 86 consecutive times, servicing robot vacuums more than 200 times in a row, packing blocks over 1,800 times consecutively, folding boxes over 200 times in succession, and packing phones more than 100 times without stopping.

Tasks trained from scratch without pretraining showed poor performance, averaging a 19% success rate. GEN-0 models fine-tuned on these tasks reached 64% success. Generalist says GEN-1 achieves production-level success rates, averaging 99%.

Generalist stated these models can respond creatively to unforeseen scenarios. In the automotive kitting example, if a washer was bumped out of proper alignment, the robot could set it down to regrasp it, partially insert it into a slit for extrinsic dexterity, or even use its other hand for bimanual in-hand regrasping.

If large deformable objects like T-shirts ended up in unexpected configurations, the model could figure out how to recover, said Generalist. "These behaviors fall well outside the training distribution and directly contribute to recovering from rare, unexpected events," the company noted.

Generalist model speeds up task completion

Generalist AI stated that GEN-1 completes tasks roughly three times faster than the current state-of-the-art (SOTA) for demonstrations. The model can dynamically react to new object physics.

For instance, GEN-1 can assemble a box in 12.1 seconds. Generalist says this is 2.8 times faster than the prior SOTA—both GEN-0 and π0 took about 34 seconds on identical boxes. GEN-1 can also pack a phone into a case in 15.5 seconds, operating at 2.8 times the speed of GEN-0.

Several components enabled these speed gains, according to Generalist. The models learn from experience and represent an evolution in inference through Harmonic Reasoning.

The company also credited its data-collection devices for providing models with access to a vast array of pretraining data from completing various other tasks at high speeds, transferring knowledge from general exposure to relevant dynamics. Generalist contrasted this with traditional teleoperation systems, which naturally produce slower, less fluid data due to lack of force feedback, latency, and visibility issues.

"Building GEN-1 was challenging—we redesigned our distributed training infrastructure to natively support petabytes of physical interaction data," said Generalist AI. The company announced that early-access partners can now gain access to the model.