Former Nvidia Executive: Why Robotics Hasn't Had Its ChatGPT Moment Yet

Why the Robotics Industry Is Still Waiting for Its 'ChatGPT Moment'

While large language models went from lab to mass-market products in just a few years, one question keeps coming up: when will robotics have its own "ChatGPT moment"?

Amulya Vishwanath's answer may not sit well with optimists. After seven years working on product marketing for Nvidia's robotics platform, she has now transitioned to venture capital. Her assessment: the robotics industry is still in its "Lego era"—actuators, robotic hands, software, and safety standards are all basic building blocks still being reassembled one piece at a time.

This judgment directly challenges a popular narrative in the current market: that all you need is a sufficiently powerful general-purpose foundation model, and robotic intelligence will take off just like language models did. Vishwanath's observation points to a more mundane reality—hardware-level engineering and mass production, along with safety certification, are the real bottlenecks holding back robots from reaching widespread adoption.

The 'Lego Era': Building Blocks Not Yet Assembled

Vishwanath uses the term "Lego era" to describe the current stage of the robotics industry, and this metaphor deserves unpacking.

Lego bricks are standardized, composable, and reusable. But today's robotics industry is nowhere near that state. Actuator designs vary from company to company, there is no unified paradigm for robotic hand degrees of freedom or sensing approaches, software stack compatibility is poor, and safety standards lack coordination across markets. This means that when developing new products, robotics teams spend a significant portion of their energy reinventing the wheel rather than building incrementally on existing foundations.

This fragmentation directly slows down iteration speed. The reason language models can evolve so quickly is largely thanks to the low marginal cost of software iteration—change one line of code and you can validate a new idea. But every hardware revision for a robot involves tooling, procurement, assembly, and testing, with cycles measured in months or even quarters.

Actuators are devices that convert energy (usually electrical) into mechanical motion, and they are the core components of robotic joints. They typically consist of a motor, reducer, encoder, and driver, and directly determine a robot's torque density, response speed, and motion precision. Currently, actuator designs in the industry are highly fragmented: harmonic reducers, planetary reducers, and cycloidal drive solutions each have their trade-offs, while quasi-direct-drive and force-control approaches remain hotly debated in collaborative robot scenarios. This fragmentation not only drives up the cost of individual robots but also makes it difficult for software algorithms from different manufacturers to migrate across hardware platforms.

Robotic hands involve another layer of complexity. Degrees of freedom (DOF) refer to the number of joints a robotic hand can independently move. The human hand has over 20 degrees of freedom, while current commercial dexterous hands typically have between 6 and 12. The higher the DOF, the greater the generality of grasping, but the difficulty of control algorithms and sensor fusion rises exponentially. Tactile sensing approaches—piezoresistive, capacitive, optical—still lack a unified standard, making grasping strategies nearly impossible to reuse across different robotic hands.

Why General Foundation Models Can't Save Robotics

The logic that "with general foundation models, robots can take off" sounds smooth, but it implies a flawed analogy.

Language models operate in the digital world, where inputs and outputs are tokens with no physical constraints. Robots, on the other hand, must complete a closed loop of perception, decision-making, and execution in the physical world, where an error in any single step can lead to task failure or even a safety incident. A grasping strategy that performs perfectly in simulation may completely fail in the real world due to changes in lighting or different surface friction coefficients.

This doesn't mean foundation models have no value for robotics. Vishwanath's view is closer to: foundation models are a necessary condition, but far from a sufficient one. Improvements in perception and decision-making capabilities can indeed unlock some new scenarios, but if actuator precision is insufficient, robotic hand dexterity is lacking, and safety certification can't be obtained, then no matter how smart the "brain" is, it cannot drive a "body" that can't reliably act.

Sim-to-real transfer is a core challenge in robot learning. In simulation environments, physics engines (such as MuJoCo and Isaac Sim) need to model physical processes like contact forces, friction, and deformation, but these models have systematic deviations from the real world—the so-called "reality gap." Lighting conditions, material reflectivity, air humidity, and even temperature changes can all cause perception models trained in simulation to fail in real environments.

Domain randomization is a common mitigation strategy: randomly varying lighting, textures, and physical parameters in the simulation during training, forcing the model to learn more robust feature representations. But this approach essentially trades computational resources for generalization ability and cannot fundamentally eliminate the gap in physical fidelity. More importantly, the safety boundaries of policies in the real world cannot be exhaustively verified through simulation—this is precisely the fundamental difference in verification paradigms between robot learning and language models.

The Underappreciated Safety Standards Problem

When discussing robotics, safety standards are often the least glamorous and least mentioned topic. But they may be the biggest hidden barrier to commercialization.

Industrial robots already have a mature safety regulatory framework, but the safety assessment frameworks for next-generation general-purpose robots—especially collaborative robots designed to share workspace with humans—are far from complete. Requirements for robot safety certification vary significantly across countries and regions, creating enormous compliance costs for robotics companies wanting global deployment.

Vishwanath lists safety standards as one of the basic building blocks, implying a judgment: before the safety certification system matures, large-scale deployment of robots will be severely constrained. This is not a problem that can be bypassed with a stronger model.

ISO 10218 is the foundational international standard for industrial robot safety, covering safety requirements for the robot itself and integrated systems. ISO/TS 15066 specifically addresses collaborative robots, establishing the principle of power and force limiting when humans and robots share a workspace, requiring that the force and pressure when a robot contacts a human body not exceed pain thresholds. However, these standards are primarily based on deterministic actions in traditional industrial scenarios. For general-purpose robots driven by learning algorithms whose behavior is difficult to fully predict, existing frameworks are hard to apply directly.

Safety certification is typically performed by third-party organizations, such as CE certification in the EU, UL certification in the US, and CR certification in China. Each certification requires extensive testing and documentation work, with timelines that can stretch to months or even over a year. For robotics companies hoping to deploy globally, this means repeating the certification process in different markets, with compliance costs potentially accounting for a significant proportion of total product development budgets. This is one reason why Vishwanath views safety standards as a basic building block.

Implications for Entrepreneurs and Investors

If the "Lego era" assessment holds, then the investment logic in robotics today needs adjustment.

Teams betting on "general-purpose robot foundation models" may be underestimating the complexity at the hardware and safety levels. Conversely, companies focused on perfecting a single building block—such as high-performance actuators, dexterous hands, or robot safety certification toolchains—may have a clearer short-term value proposition.

Another noteworthy signal is that Vishwanath's move from a product marketing role at Nvidia to VC itself reflects the industry's demand for talent who truly understand the challenges of robot deployment. At this stage, the value of industry experience and engineering judgment may exceed pure technical storytelling ability.

Robotics won't have an overnight "moment"—it will be more like a process of building blocks gradually becoming standardized, assembled, and ultimately forming a reusable system. Patience and infrastructure building are the scarcest resources in this sector right now.

One-line Summary

Former Nvidia robotics executive Vishwanath believes robotics is still in its "Lego era," with basic components and safety standards as the bottleneck.

Paragraph Summary

Why is the robotics industry still waiting for its "ChatGPT moment"? Amulya Vishwanath, former Nvidia robotics platform executive and now venture capitalist, offers a sober assessment: the robotics industry remains in a "Lego era," where actuators, robotic hands, software stacks, and safety standards have yet to be standardized. General foundation models are a necessary but not sufficient condition for robotics—the perception-decision-execution loop in the physical world is far more complex than token processing in the digital world. The sim-to-real gap, the high cost of hardware iteration, and the fragmented safety certification system together form hidden barriers to commercialization. The article further notes that teams betting on general-purpose robot foundation models may be underestimating hardware complexity, while companies focused on basic components like actuators, dexterous hands, and safety certification toolchains actually have clearer short-term value propositions.