Nomura report: The ChatGPT moment for global robotics has not yet arrived

Recently, the humanoid robot sector has continued to receive favorable catalysts, with Unitree Robotics at one point raising prices by 20CM, triggering heightened market attention to the robot talent industry.

A few days ago, Nomura released a research report on global robotics, whose views are relatively sober and offer a more important judgment: the robot's true "ChatGPT moment" will not be announced by a single launch event, but will be quietly confirmed by a set of operating metrics and a data flywheel.

The report likens current robot technology to the AlexNet moment of 2012—deep learning has already proven its value, but the robot brain has not yet reached the GPT-2 level; new models in December to January may be closer to GPT-2, yet still do not meet the conditions for mass production.

In other words, the robot industry is standing on the eve of an explosion, but the explosion is not a single-point event; it is an endurance race shifting from a model competition to a data flywheel, hardware consistency, and operational validation.

From AlexNet to GPT-2: The True Coordinates of the Robot Brain

The core analogy of the report is that robots are retracing the old path of large language models, only starting earlier.

LLMs went through three stages—massive pretraining, instruction fine-tuning, and reinforcement learning through trial and error—and robots will do the same. Nvidia scientists judge that robot training will likewise move from broad learning to task-specific fine-tuning, and then improve continuously through trial and error.

But today's robot brain has not yet reached the GPT-2 stage:

Sprouts already exist in the lab, yet no model is capable enough for real work.

General Artificial Intelligence's 1.5 model can learn a task from just 3 to 12 seconds of demonstration, with a 59% success rate over ten tests without additional training, rising to 83% after supplementary training with only five minutes of data; Skild AI's S1 can learn to make coffee by watching human videos; Ant Group Robbyant's Zero-WAM achieves a 47.0% success rate in simulated environments across seven unseen tasks, 29.5 percentage points higher than the best of its kind.

These results indeed show signals of GPT-3-style few-shot learning, but all come from internal company tests, with short tasks or in simulated environments, and Skild AI's CEO also said outright that S1 is not yet ready to enter homes.

Therefore, the report still positions the robot brain at the GPT-2 level or lower.

Where is a real breakthrough more likely to come from? Nomura's judgment is large AI labs and Tesla, rather than pure robot-brain startups.

Labs possess compute, talent, and general models, while Tesla has a closed data loop, factory testing environments, and a robot fleet already in use.

Google's Gemini Robotics 2, released in July 2026, achieves a success rate as high as 91.3% on complex multi-step tasks, with full-body control about a year ahead of Chinese competitors;

OpenAI is expected to release a model designed specifically for physical AI in December to January, and Altman expects that in 2027 people will see "impressive demonstrations" of OpenAI robots;

Wang Xingxing, meanwhile, revealed that Unitree has already tested GPT-6 on robots with very ideal results, and believes that next year GPT-7 is expected to break through the embodied AI bottleneck.

English text is being prepared.

Three-Line Race: VLA, World Models, and Simulation Reinforcement Learning

The technical routes for the robot brain are diverging into three: VLA, world models, and reinforcement learning in simulation.

VLA observes scenes, receives instructions, and outputs actions, usually as paths that a lower-level controller can follow; onboard VLA models have about 2 to 7B parameters and require about 30,000 to 40,000 yuan of chips and storage.

World models, by contrast, predict what will happen after the robot acts?

According to Fei-Fei Li's framework, it includes a renderer, a simulator, and a planner; World Labs' Atlas reaches 100 to 200B parameters, and a world model that generates full video requires 3 to 10 times more compute, while a streamlined version running on compressed internal representations requires only 30% to 100% more.

It is expected that by 2027, world models will run in the cloud, responsible for training robots and generating data, while compact VLA and RL models run on the robot body.

Simulation reinforcement learning, meanwhile, trains actions through trial and error inside physics simulators; Unitree's walking robots are the most famous example, and skills trained only in simulation have already worked on real machines.

These three routes are not mutually exclusive; they are competing to become the robot's "brain."

The report notes that early VLA models such as π0 and GROOT spent most of their capacity on language, so they know what an object is but not how it moves; WAM instead uses video generation as a physics engine, and predicting video frame by frame can teach a model gravity, buoyancy, and light, predicting in one step what will happen next and what the robot should do, thereby handling new scenes without additional training.

Dream Dojo research goes further, replacing the physics engine with a video model so that robots can practice in many simulated worlds at once.

The SuperMimic paper demonstrates a transformation in training methods: running for three hours on just one GPU, the model's performance surpassed a baseline model trained for seven days on 128 GPUs, reducing compute resources by more than 1,000 times; in 129,514 action tests, the old method failed 179 times, while the new model failed 25 times in three hours and achieved zero failures after 9.8 hours on two GPUs.

This suggests that we should train robots within learned models rather than in the original simulator, and use more motion data to train a single controller.

But the research report also emphasizes that pretraining is only half of the process; in order to keep improving, models need data generated by robots working in the field.

Data flywheel: the 1 million-hour threshold and the video scaling law

Data is precisely the bottleneck constraining robot development.

Language models can learn from text across the entire internet, but robots lack a similar data source.

Teleoperation lets humans guide robots to collect data, which is both expensive and slow, and the data is tied to a specific robot model—once a robot changes from legged to wheeled, the data can no longer be used.

The research report puts the threshold at about 1 million hours of data, roughly equivalent to the time accumulated over a person's lifetime; currently, the usable data collected in a unified format is only about 100,000 hours, and leading companies have only a few hundred thousand hours of high-quality data, while the data collected by Tesla's Optimus has just surpassed 500,000 hours.

Starting training requires 10 million hours, and the long term requires more than 100 million hours, while the industry has so far accumulated only a few million hours.

More critically, much of the data is invalid: Unitree can use only about a quarter of its recorded time, with only one hour valid out of every four hours recorded.

The market structure of data collection also confirms where the bottleneck lies.

Nomura estimates that China's data collection market will be about 4 billion to 5 billion yuan in 2026, of which physical robot teleoperation is about 2.4 billion yuan, first-person video and UMI about 1.3 billion yuan, simulation about 600 million yuan, and failure recovery about 300 million yuan.

In terms of cost, teleoperation is about 700 yuan per hour, UMI about 100 to 300 yuan, first-person video about 300 to 350 yuan, and simulation about 120 yuan. Teleoperation data is the most expensive and tied to robot models; first-person video and UMI are cheaper, and video, as camera footage, can be reused across different generations of robots; simulation is the cheapest, but accounts for no more than 20% of data in real-world perception tasks.

This is precisely why the industry is turning to video: a typical data mix is already 60% first-person video, 30% UMI, and 10% real robot data.

NVIDIA's EgoScale first trains on more than 21,000 hours of everyday first-person video, then fine-tunes with a small amount of teleoperation and motion-capture data; for hand manipulation, more pretraining compute brings stable and predictable performance gains, consistent with the scaling laws that drove the development of LLMs.

If human video becomes the main data source, then whoever has the best data-processing pipeline and the strongest compute will gain the greatest advantage. This further highlights the importance of large AI labs and limits the long-term value of teleoperation data tied to specific robot designs.

But the real ChatGPT moment lies in the positive feedback loop starting to run: robots at work generate data, the data lets models improve, and better models are applied more widely.

Tesla plans to train by the end of 2026 through about 10,000 robots at its own facilities, collecting 1 million hours of valid data; from the second quarter to the third quarter of 2026, the data cost for Optimus fell by about 18.1%.

By comparison, China's deployment scale is much smaller: in 2026, the seven largest express delivery companies are expected to purchase fewer than 500 humanoid robots, and in 2027 only about 1,000.

The research report therefore concludes that the positive feedback loop will start first at a company that owns both robots and models.

OpenAI purchased about 10 million hours of data between September 2025 and August-September 2026, accounting for about 70% of U.S. supply; DeepMind procured about 20%, equivalent to one-third of OpenAI's purchases.

Independent data collectors still have a collection window of about three years, but as Chinese data cannot be exported and AI companies rely on their own collection, the number of data suppliers is expected to decline by the end of 2027.

In other words, the threshold for the data flywheel is not "whether there is data," but "whether valid data can be continuously obtained at low cost and iterated in a closed loop."

Dexterous hands and consistent manufacturing: the iPhone moment on the hardware side

Once the model path is determined, humanoid robot design will gradually converge into a few standardized forms, which the research report calls the "iPhone moment" in the hardware field. After that, the hardest part will be how to manufacture every device in a unified or similar way.

A humanoid robot has more than 20 joints, and if there are subtle differences in behavior between different robots, a control policy trained on one robot may apply to only a few other robots.

Therefore, the ultimate winners will be companies that can iterate quickly and strictly control costs, such as Unitree Robotics, automakers, and key component suppliers.

In terms of mass-production cost, the research report estimates that a 1.7-meter-tall robot costs between 150,000 and 200,000 yuan, and a 1.3-meter-tall robot costs between 30,000 and 40,000 yuan; actuators—motor and gearbox components—account for 50% to 75% of component costs.

Unitree currently designs about 80% to 90% of its BOM in-house, manufactures about 50% in-house, and uses third-party modules for 10% to 15%; as production increases, it is expected to increase outsourcing, both to maintain low costs and to ensure quality.

In the past, many robots used grippers simply because there was too little data available to train hands; more powerful foundation models can transfer skills learned on two-finger or three-finger grippers to five-fingered hands, thereby achieving greater grasping capability and lowering prices.

Nomura expects dexterous hand shipments to grow 80% to 120% year over year in 2026, with high-degrees-of-freedom technology leading this trend; this year, 30% to 50% of factory fixtures will be replaced.

In terms of price, the average ASPs of different technology routes such as linkage, tendon-driven, and direct-drive are all declining, but shipments of high-degrees-of-freedom hands are growing faster. The explosion on the hardware side is not simply "replacing grippers with human hands," but a process in which, as AI capabilities improve, hardware value shifts from grippers to dexterous hands, actuators, and key components.

Five operational metrics: the hard threshold for the breakout moment

Nomura proposed five tests to judge whether robotics technology has entered its "ChatGPT moment," and currently robots fall far short of expectations on all of these tests.

First, how long a robot can continue working before requiring human intervention: currently the best-performing robots can only run for a few hours, while factories require continuous operation for a week without manual intervention.

Second, fine manual manipulation capability: the Apollo 2 robot equipped with the Sharpa Wave hand—Gemini Robotics 2—has only a 44% probability of tying a garbage bag tightly and only a 32% probability of sweeping garbage into a dustpan.

Third, the speed at which models adapt to new robot bodies: Google needs about seven to eight hours, while more technologically advanced Chinese manufacturers need about two weeks.

Fourth, the number of demonstrations required to complete a new task: in laboratory settings this has already fallen to one or two.

Fifth, operational precision in factory environments: current industry test results are only about 60%, while the industry requires more than 98%, and Nomura expects it will take three to five years to close this gap. The research report considers the first and fifth items the most important, because they determine whether customers are willing to pay for robots out of operating budgets rather than trial budgets.

These five metrics constitute the hard threshold for the "true breakout moment." Model releases, demo videos, and few-shot learning cases can create optimism, but only continuous operation for a week, factory precision above 98%, rapid adaptation to new bodies, and a low number of demonstrations can move robots from trial budgets into operating budgets.

For this reason, Nomura positions the current stage as an acceleration period after the "AlexNet moment" and before GPT-2, rather than the ChatGPT moment itself. The true breakout moment is when the data flywheel, model capabilities, and hardware consistency all cross the critical point at the same time.

Conclusion

The robotics industry is ushering in a dense round of catalysts: OpenAI's physical AI model, new models from ByteDance and DeepSeek, OpenAI's humanoid robot demonstration, and GPT-7, expected to launch in 2027.

But Nomura's research report analysis shows that these events are more likely to push the robot brain from below GPT-2 to near GPT-2, rather than directly heralding the ChatGPT moment.

The true breakout moment requires three conditions to hold simultaneously: a large AI lab or Tesla-style compute and data closed loop, a data flywheel centered on video and world models, and hardware-consistent manufacturing brought about by dexterous hands and standardized bodies. The ultimate winner will be the companies that simultaneously possess strong computing power, a data closed loop, and robots deployed in the field.

The true breakout moment for robots will not be a day when a certain model suddenly astonishes the world, but rather the quiet start of a flywheel in which robots at work begin to continuously generate effective data, models are optimized accordingly, and more robots are therefore deployed.

That moment will not come with drums and gongs, but it will leave irreversible traces in one operational metric after another.

This article is from the WeChat public account "Gelonghui APP" (ID: hkguruclub), author: Godzilla, published by 36Kr with authorization.