Author: RoboStrategy Research Team
Compiled by Deep潮 TechFlow
Shenchao Summary: Within just two months of its release, ChatGPT garnered hundreds of millions of users, igniting humanity’s fervent imagination around artificial intelligence—and prompting a long-standing question for the entire embodied intelligence and robotics sector: When will robots have their own ChatGPT moment? In this in-depth research report, the RoboStrategy team offers a starkly contrasting perspective: Robots will never experience a so-called ChatGPT moment.
The purely digital output of the software world inherently allows for trial and error and manual post-hoc fine-tuning, while inexpensive cloud computing infrastructure drives marginal distribution costs close to zero; in contrast, interactions in the physical world are constrained by heavy capital hardware costs, irreversible error consequences, and extremely stringent physical security boundaries.
Every idle moment of a robot is a tax on asset returns, and it must scale the cliff of 99.9% or higher reliability to earn its ticket to commercial deployment. But this does not mean embodied intelligence is falling behind; on the contrary, from Waymo’s scaled deployment of autonomous fleets to the vertical integration of embodied models in welding, food service, and logistics, physical AI is leveraging multimodal foundation models and scaling laws to follow a penetration curve similar to autonomous driving: gradual, then sudden—trading reliability for distribution, while the true industrial explosion has already quietly begun in the deep waters.
Back to November 30, 2022. OpenAI released ChatGPT.
In 5 days: Users reach 1 million.
Two months later: User base exceeds 100 million.
It was a starting gun heard around the world. The unprecedented speed of user adoption was only part of this “moment.” What truly mattered was that the global collective consciousness simultaneously underwent a paradigm shift in thinking: AI was no longer just a scientific fantasy confined to laboratories. It was clearly real—and poised to ignite the next technological revolution. Corporate executives across industries elevated AI planning to their top priority. Investors raced to position themselves in potential winners or urgently exit companies about to be disrupted. Entrepreneurs launched new ventures in semiconductors, biotechnology, and robotics. AI would inevitably sweep through every sector. Ordinary people around the world also began using AI chatbots daily in their everyday lives.
Since then, everyone in the robotics field has been asked the same question: When will our robot ChatGPT moment arrive?
The robotics field will never experience any so-called ChatGPT moment.
Figure 1: Jensen Huang stands on stage alongside representative humanoid robots from various companies, discussing whether the "ChatGPT moment" for robotics has arrived.
Intelligent robots have already arrived among us and will be adopted on a large scale, just like any other transformative technology in the past—but this will not happen within weeks; it will take years.
The internet, smartphones, and most recently, autonomous vehicles, all confirm this pattern. Anyone still waiting for the “robot ChatGPT moment” is overlooking a key fact: the starting gun was fired years ago, and the first wave of real-world, large-scale deployments has already begun this year. Companies from UPS to BMW to Samsung are pouring vast resources into automating operations with robots. Top talent is flooding into the field, building a new ecosystem of Physical AI companies, and capital is following suit—though it still pales in comparison to the massive sums that flowed into pure software AI over the past several years.
So why is it impossible to see a ChatGPT-like explosion—from zero to 100 million users in two months—on robots? Beyond the physical manufacturing constraints of producing 100 million robots, the capabilities required for robots to be deployed differ fundamentally from those of general-purpose AI models. We often forget how poor ChatGPT was when it first launched: it frequently hallucinated, provided answers riddled with errors, and couldn’t even answer questions that middle school students could solve. Yet in certain large-scale use cases, these flaws didn’t matter much. In tasks like copywriting, text summarization, and basic research, it still demonstrated practical value. It didn’t replace entire workflows, but it saved ordinary people small amounts of time in their daily work. Fundamentally, it was a mediocre but useful tool—few were willing to pay a high price for it, but many were happy to pay a small fee or use it for free to boost productivity. More importantly, since it ran in the cloud, its marginal delivery cost was extremely low. No one needed to buy or host their own GPU hardware; everyone shared the same centralized cloud infrastructure. When the marginal cost of delivering a service is negligible, the barrier to making it “worth using” becomes incredibly low.
However, the bot faces two core pain points that chatbots have never encountered before.
First, there is the “idle-robot tax.” Whether a robot is actively working or standing idle, the buyer must continue paying for it—either as a capital expenditure (CapEx) with an upfront full payment, or as an operating expenditure (OpEx) through leasing or Robot-as-a-Service (RaaS). A robot used for only one hour per week must generate extraordinarily high value during that single hour, because the cost of relocating and retraining it for other tasks is prohibitively high. As physical AI enables general-purpose robots to handle multiple tasks and move with human-like flexibility, I expect this dynamic to change. But for now, every idle hour of a robot represents a heavy tax burden.
Second, there is the “99.9% wall.” The criteria for what constitutes a “passing standard” differ drastically between knowledge work and physical labor. A mediocre, rough draft can still save time, because humans can retain the essentials and refine the flaws before finalizing. Copy, presentations, and even legal contracts are all subject to repeated revision.
Physical labor is far less forgiving. A poor weld can completely ruin an entire metal component; an incorrect dosage of medication can lead to irreversible, catastrophic consequences.
Figure 2: Classification matrix of knowledge work versus physical work along the dimensions of error recoverability and quality tolerance. Physical work is predominantly concentrated in the high-risk zone characterized by low recoverability and extremely narrow tolerance.
The nature of knowledge work allows for corrections and revisions to be made to AI outputs before they are delivered as final products. In contrast, while humans can also intervene in physical labor, the cost of recovering from a single physical error is often far greater than editing a paragraph. Physical labor typically has very low tolerance for errors, meaning it demands extremely high levels of reliability. For robots performing physical tasks to have practical value, they must consistently produce high-quality results—not just achieving 99% success rates, but often requiring reliability above 99.9%. This creates a much higher barrier to scalable deployment. Beyond occasional successful demonstrations, robots must complete tasks with sufficient reliability at a low enough cost to truly convince customers to integrate them into actual operational production.
Figure 3: Comparison of success rates required for practical thresholds between knowledge work and physical labor. Knowledge work typically creates value with success rates between 50% and 90%, while physical operations often demand absolute reliability of 99.9% to 99.9999%.
There is a vast difference in the quality thresholds for work deemed "practically useful" between knowledge work and physical labor.
In the years since the release of ChatGPT, the field of robotics has made tremendous leaps forward, with some of this progress driven by spillover from external technologies. General AI models and video generation models have jointly accelerated robotics research and development, forming the foundation of a robot’s core intelligence.
Many cutting-edge robotic models today employ a dual-system architecture, inspired by psychological classifications of human thought processes. System 2 is the slow, deliberate thinker: a specially adapted general-purpose AI large model responsible for understanding environmental contexts, reasoning, and planning. System 1 is the fast, instinctive executor: an action execution model (in some architectures derived from video prediction models) responsible for translating these plans into high-frequency motor control commands in real time.
This means that every technological breakthrough achieved by OpenAI, Anthropic, Google DeepMind, NVIDIA, Alibaba, and leading university laboratories is directly or indirectly contributing to the advancement of physical intelligence. The table below provides a comprehensive comparison between the core capabilities of cutting-edge robotic models at their initial release and those of ChatGPT today.
Figure 4: Comprehensive evaluation comparing frontier robotic models (such as Dyna 2.1 and Rhoda DVA) with ChatGPT (GPT-3.5) at launch and current general-purpose models across dimensions of learning and memory, autonomous execution, and reliability.
In my view, cutting-edge robotic models have already surpassed the GPT-3.5 model that powered ChatGPT at its initial launch across several key dimensions. Embodied models from companies such as Dyna Robotics and Rhoda AI can now handle long workflow processes, perform multi-step reasoning, maintain extended context states, and actively detect their own errors while autonomously retrying and correcting them.
Meanwhile, we have witnessed the first physical AI companies transition from laboratories and prototype testing into large-scale real-world deployment. In fact, Waymo has been pursuing this path in the field of autonomous driving for several years.
Over the past few years, the internet has been flooded with impressive videos showcasing general-purpose robots successfully completing tasks. But the core issue has always been that outsiders had no way of knowing whether these polished short clips were carefully edited down from dozens—or even hundreds—of failed attempts (which, in fact, they often were). Figure AI’s live-streamed 200+ hours of uninterrupted footage of its humanoid robot sorting packages represents a “mini ChatGPT moment” for robotics: for the first time in history, there is real-time, undeniable evidence that general-purpose machines can perform complex physical labor as reliably as humans.
Path Robotics has now widely deployed its industrial welding robots in the operations of leading customers across energy, defense, and shipbuilding sectors, and earlier this year announced a substantial contract with defense giant HII worth up to $600 million (with a potential maximum of $900 million). Dyna has similarly rolled out its automation robots across networks of large restaurant chains, hotels, and commercial laundries, reliably managing a variety of complex processes.
Figure 5: Compilation of recent reports from major global business media and institutions on the large-scale deployment of Figure AI, Standard Bots, Dyna Robotics, and Path Robotics in real-world factories and commercial networks.
How is all of this achieved? The answer lies in specialized models and targeted fine-tuning. The models that navigate Waymo’s autonomous vehicles or enable precise welding for Path are not suited for organizing bins. However, the scale of these vertical markets is large enough to justify training dedicated models for mastering a single skill and ensuring they operate reliably across a wide range of complex real-world scenarios.
Embodied AI companies such as Dyna and Rhoda typically begin with general-purpose robotic base models, collect real-world physical data for specific workflows, and continuously fine-tune until the models achieve the stringent reliability standards required for industrial applications. They prioritize entering the most widely used and pain-point-intensive general processes, gradually expanding the scope of physical data collection as subsequent funding rounds and capital accumulation progress. This approach mirrors that of general AI labs, which initially focused on solving problems in software engineering and coding.
In the past, achieving commercial-grade reliability for robots in single tasks often took months. However, as underlying foundation models have become increasingly sophisticated, leading companies have found this timeline is rapidly shrinking to just weeks. Earlier this year, several cutting-edge teams published new research providing conclusive evidence that robotic models also follow scaling laws. In other words, they assert that increasing data volume and computational resources leads to a qualitative leap in the generalization capability of embodied models, significantly reducing the amount of fine-tuning data required for each new skill. With each reduction in post-training costs, the friction to deployment decreases, extending the list of tasks general-purpose robots can perform.
So, what will the actual path to widespread adoption of robots look like?
Figure 6: Growth trajectory and forecast for autonomous vehicles on the road from 2021 to 2026. Starting from an initial scale of 5 million vehicles, the number is expected to surpass 16 million by 2026, reflecting a typical penetration curve of accelerating adoption of physical intelligence.
You can view autonomous vehicles—considered a crucial subcategory of physical AI—as an excellent reference model. Over the past few years, autonomous driving algorithms have undergone a qualitative leap, with the global fleet of autonomous vehicles surpassing 10 million units. At the same time, autonomous vehicles face unique challenges never encountered by other industrial robots: their reliability requirements are two to three orders of magnitude higher, as errors can result in loss of life; they must navigate the lengthy and complex process of obtaining national safety regulations and local operating permits from cities and states; additionally, autonomous taxi fleets like Waymo and Cybercab must establish dedicated, capital-intensive physical operational networks in each city, including dispatch centers, charging stations, vehicle cleaning and maintenance, remote safety assistance, and 24/7 roadside rescue teams.
Unlike ChatGPT’s consumer-focused breakout or Waymo’s penetration into ride-hailing, I anticipate that robots will achieve meaningful, large-scale commercial deployment in business and industrial sectors before truly exploding in the consumer market. I firmly believe that one day, home service robots will be as ubiquitous as cars are today; however, in the near term, training robots to achieve absolute reliability in more structured environments—such as industrial laundry, machine tool loading/unloading and inspection, and warehouse material handling—remains a more controllable engineering challenge. The gears of scaling are already turning, and the technical limits of physical intelligence are being reset every week.
Therefore, the “ChatGPT moment” may have been a flawed analogy from the start. In the world of ChatGPT, channel distribution preceded system reliability; in robotics, reliability must be established before any scale of distribution, and the engineering pathways to achieve this have now been clearly validated. We stand at the historical starting point of adoption curves similar to those seen in autonomous driving across industries. History may not simply repeat itself, but it always rhymes.
Disclosure: The RoboStrategy portfolio may include companies discussed in this article.