Generalist AI Launches GEN-1.5, Seen as the GPT-3 Moment for Robotics

On August 19, Unitree Technologies listed on the STAR Market, opening at RMB 1,100 per share with a market capitalization of approximately RMB 444.9 billion.

On the same day, Jensen Huang's daughter, Huang Minshan, arrived in Beijing to visit the WRC World Robot Conference.

The very next day, at Beijing time凌晨, Silicon Valley robotics company Generalist AI released its next-generation foundational model, GEN-1.5.

Two months ago, Stanford professor Fei-Fei Li invested $400 million in Generalist AI as an individual investor; other angel investors in the same round included Lin Bin, co-founder of Xiaomi, and Yuan Zheng, founder of Zoom. NVIDIA is also a shareholder. Smart money and smart people are converging on the same direction.

GEN-1.5 provided their rationale for investment: showing robots a 3- to 12-second action demonstration, without training or fine-tuning, and achieving an average success rate of 59% across 10 manipulation tasks.

Previously, teaching a robot to twist off a bottle cap required three months of programming. Now, all it takes is a three-second demonstration and zero lines of code.

Several leading researchers have compared this moment to the release of GPT-3 in 2020, believing that embodied intelligence has reached its own GPT-3 moment.

What surprised researchers most about GEN-1.5 is not its ability to imitate, but its capacity for improvisation.

A generalist AI conducted an experiment: using just five minutes of human demonstration data (with only one gradient step of fine-tuning), it taught a robot to sweep blocks into a bowl with a brush, then swap tools.

Give it a banana, and it uses the banana like a brush to sweep the blocks into the bowl—this isn’t surprising, as the banana’s shape and contact strategy resemble those of a brush.

Give it a dustpan, and it abandons the "sweeping" strategy, using its other hand to push the blocks onto the dustpan, lift the dustpan, and pour the blocks into the bowl.

“Sweeping” and “picking up and dumping” are two completely different contact sequences, and this set of movements was not taught in the training data.

The generalist AI searched through approximately 1.89 million pre-trained scenarios using nearest-neighbor language search but found no similar usage of a shovel.

No one told the model to do this at any stage.

Similar improvised behaviors repeatedly occurred during testing: the bowl was covered with a piece of paper, and the model would lift the paper, place the blocks, sometimes even replace the paper—scenarios not present in the training data.

A LEGO brick is stuck to the tip of a mechanical finger, and the model uses its other hand to knock it off.

The training data only demonstrated rotating the jar lid with one hand, but the model spontaneously switched to using both hands during some attempts, adopting a completely different grasping and rotating strategy.

A model trained only to place one block into a bowl begins spontaneously sorting multiple blocks by color.

These capabilities share a common source: large-scale pretraining.

GEN-1.5 has been continuously pre-trained for over eight months, spanning three training phases.

The validation curve published by Generalist AI shows that the model's "next action prediction error" continues to decline, with no signs of convergence to date.

In terms of large language model terminology, this is the Scaling Law for embodied intelligence: as the volume of physical interaction data increases and training duration extends, the model's generalization capability continues to improve.

An counterintuitive finding makes this insight clearer: the fewer fine-tuning gradient steps, the stronger the improvisational ability.

The explanation for Generalist AI is that light fine-tuning brings the model closer to the broad set of behaviors accumulated during pre-training, preserving more of the accessible "physical experience."

Ten gradient steps changed the model parameters by only 0.15%.

The generalist AI says: This is almost like reminding the model of something it already almost knows.

GEN-1.5 also bridges two long-standing gaps in the field of robotics.

Record a demonstration in the simulator (even though the pretraining data contains no simulated data), feed it into the context window, have the real robot execute the task directly, and achieve generalization across different robotic hands and new object positions.

In some tests, a person demonstrated movements directly in front of the robot's camera using their own hands, and the robot subsequently replicated the task using its mechanical arms.

The generalist AI indicates that none of these capabilities were the result of deliberate design: there were no architectural modifications specifically aimed at promoting in-context learning, no meta-learning loops, and no auxiliary training objectives designed to encourage improvisation.

These capabilities emerge spontaneously from pretraining on large-scale physical data.

Other research teams have previously demonstrated similar few-shot learning capabilities, but only for a limited range of task types; the novelty of GEN-1.5 lies in its broad coverage. All results were self-reported by the Generalist AI and have not yet been independently verified.

The task itself is a simple short-range operation (tightening a bottle cap, zipping a zipper), still far from the long-range, complex tasks found in real-world scenarios.

GPT-3 was released in June 2020, and it took two and a half years (yes, two and a half) from GPT-3 to ChatGPT.

GEN-1.5 is currently in a position similar to GPT-3 in 2020: capable but rough, with clear direction, simple tasks, and a clear scaling trend.

Generalist AI wrote a thought-provoking passage on the official blog:

After surpassing a certain pretraining threshold, the cost of adapting to a new task becomes negligible.

Emergent in-context learning, seconds of data, or a single gradient step combined with a one-minute demonstration is no longer traditional task-specific training; it is more like reminding the model of something it almost already knows.

If the scaling curve of embodied intelligence shows no signs of convergence (as Generalist AI says they haven’t seen any), then the core resource for the next round of competition will become who has enough physical interaction data to feed this engine.

This is exactly like the script of the data competition that large language models entered in 2021.

In June 2020, OpenAI released GPT-3.

This model did something no one had achieved before: without retraining, it could perform new language tasks simply by providing a few examples in the chat box—given examples like “red → apple, yellow → banana,” and then asked “green → ?,” it could answer “watermelon.”

This capability is called in-context learning, marking the turning point where large language models transition from being "specialized tools" to becoming "general-purpose platforms."

GEN-1.5 brings the same concept into the physical world.

Generalist AI refers to it as Physical Prompting.

The method is straightforward: feed a 30-second context window of the model with a real-world action demonstration (including sensor data and motion trajectories), leaving the remaining capacity to receive real-time environmental observations.

The model attempts to execute immediately, with no training steps involved.

Here is an essential concept to understand for GEN-1.5: the gradient step.

In traditional approaches, teaching a robot a new task requires tens of thousands of gradient descent steps, each involving a fine-tuning of the model’s internal parameters—a process that typically demands vast amounts of data and computational power.

The one-shot mode of GEN-1.5 skips all gradient steps, leaving the model parameters completely unchanged.

The Generalist AI evaluated GEN-1.5 on 10 distinct operational tasks: opening a glass jar lid, zipping a pencil case, retrieving money from a wallet, folding paper, stacking cups, flipping a smartphone, sweeping blocks with a brush, opening a book cover, peeling off a vacuum seal, and cleaning up debris.

Results are divided into two groups:

One-shot (view once, zero gradient steps): 59% average success rate across 10 tasks (standard deviation ±10%).

Few-shot (few-shot, 5 minutes of data, approximately 50 demonstrations, 10 gradient steps): average success rate of 83% (standard deviation ±9%).

For comparison, in 2020, GPT-3 achieved approximately 45% on one-shot language tasks and approximately 65% on few-shot (around 100 examples) tasks.

The structure of the two sets of numbers is highly similar.

The meaning of this set of numbers can be understood as follows: Before the emergence of large language models, getting AI to perform a new language task—such as translating an unfamiliar format—required collecting specialized data and training a new model, a process that took months.

After GPT-3, this task became "writing a few examples in the input box," reducing the time cost from months to seconds.

GEN-1.5 achieved the same compression in the domain of physical operations.

Previously, teaching a robot to twist off bottle caps required engineers to program for months or to fine-tune parameters using tens of thousands of training examples.

Now, a 3- to 12-second demo is sufficient.

The Generalist AI wrote on the official blog: This changes two things—the speed at which robots become useful (from months to seconds) and who can use them (from experts to anyone).

The release of GEN-1.5 occurred during the most intensive week for the embodied intelligence industry.

Starting on August 19, the WRC World Robot Conference opened in Yizhuang, Beijing, featuring over 300 companies, more than 2,000 exhibits, and more than 150 global debut products.

On August 22, three days from now, the second World Humanoid Robot Games will kick off at the National Speed Skating Oval, "Ice Ribbon," with 666 teams competing in 1,301 matches across 51 events, using 2,056 robots—a 138% increase in team count compared to the first edition.

The sports event features 21 scenario-based competitions for the first time, requiring humanoid robots to complete long-distance tasks in real-world environments such as factories, hotels, and home services, with the theme being "Robot Deployment Testing."

Yushu Technology, which just rang the bell, shipped over 5,500 humanoid robots in 2025, ranking first globally, with revenue of RMB 1.7 billion and a gross margin of 60%. DeepSeek Also participated in the strategic placement.

However, Yuyu’s prospectus contains a set of contrasting figures: revenue growth in the first quarter of 2026 declined from 332.64% year-over-year to 68.49%, and non-GAAP net profit fell by 52.55% year-over-year. The primary reason for the slowing growth is a significant increase in R&D investment.

What Unitree is chasing is precisely what it does not yet have: an embodied intelligence large model.

"New Position Pro" previously wrote in an analysis ("In Unitree's IPO Prospectus Lies Anxiety Over the 'Shelf Life' of an Engineering Miracle") about Unitree's structural challenges: Unitree has produced the world's best-selling robotic bodies, but its "brain" is absent.

2026 is dubbed the "big year of IPOs" for embodied AI, with over 20 companies having confirmed IPO plans, but the driving force behind this IPO wave is largely financial pressure.

Many competitors rely on successive rounds of funding to stay afloat, with their R&D pace dictated by funding windows.

Image source: LatePost's "The Money Game of Embodied Intelligence"

Wang Xingxing, founder of Unitree, publicly stated at WRC that the main bottleneck limiting the development of humanoid robots is embodied intelligence large models: “We expect the robot equivalent of ChatGPT to emerge within two to three years at the earliest, or five to ten years at the latest.”

The release of GEN-1.5 can be regarded as the first empirical response to this assessment.

The Scaling Law exists in embodied intelligence, where large-scale pre-training enables adaptation to new tasks with marginal costs approaching zero.

This conclusion has direct industry implications for body manufacturers like Unitree: once a general-purpose model at the brain level matures, the value of the body will depend on how quickly it can integrate with that brain, not just on hardware specifications or shipment volume.

The team background of Generalist AI is highly aligned with this technological approach.

Co-founders Pete Florence and Andy Zeng come from Google DeepMind’s robotics team, and Andrew Barry comes from Boston Dynamics.

From left to right: Pete Florence, Andrew Barry, Andy Zeng

The company completed a $400 million funding round in June 2026, led by Radical Ventures, with participation from NVIDIA and Bezos Expeditions, resulting in a post-money valuation of $2 billion.

Generalist AI is currently negotiating a new funding round at a $3 billion valuation.

This release has generated a strong reaction in the robotics research community.

Co-founder Pete Florence wrote on X:

Since I began researching robotic foundation models, broad one-shot contextual learning has been my clearest goal.

Now I see nearly a decade of imagination come to life in the real world.

Florence mentioned that he and Andy Zeng discussed in graduate school a "Ctrl-C-Ctrl-V"-like capability: seeing an action and having a robot replicate it—but realizing it was extremely difficult because "the world you want to paste into is never the same as the world you copied from."

Carnegie Mellon University robotics researcher Chris Paxton called GEN-1.5 "possibly the real GPT moment" on X, writing:

The holy grail of operational learning is undoubtedly one-shot learning.

https://x.com/chris_j_paxton/status/2090270734816092334

https://x.com/chris_j_paxton/status/2090210797125611972

Jimmy Yang, an embodied intelligence researcher at Northeastern University in the United States, shared and said:

As an embodied AI researcher, this is a truly special moment for the field.

https://x.com/JimmyTYYang1/status/2090202923632366073

Jim Fan, Director of Robotics at NVIDIA, provides an insightful technical perspective.

He pointed out two key factors behind the emergent capabilities of GEN-1.5:

First, naturally occurring symmetric repetitive action patterns in training data, such as repeatedly screwing in screws when assembling furniture, where the second screw serves as a "contextual learning sample" for the first;

Second, the recovery actions after human errors in the data—such as picking up something that was dropped and continuing—create a complete "failure-recovery-continue" arc, enabling the model to naturally demonstrate error-correction capabilities during testing.

Jim Fan also noted that the UMI data collection method used by Generalist AI—where humans directly wear robotic grippers to operate—preserves human "physical intuition," whereas traditional teleoperation through VR devices causes significant loss of this intuition.

https://x.com/DrJimFan/status/2090465981240086992

Of course, rational cold water also exists.

Jim Fan wrote in the comments of the same tweet:

The demonstration is still too simplistic to draw any conclusions.

https://x.com/DrJimFan/status/2090469108764823587

This article is from the WeChat public account "New Intelligence Yuan," authored by ASI Revelation.