Physical AI Is More Than Robots: A New Investment Perspective from a Former Nvidia Executive

After seven years deep in product marketing for Nvidia's robotics platform, Amulya Vishwanath has now pivoted to becoming an investor in the physical AI space. She offers a thought-provoking perspective: robotics is currently in a "Lego era," where fundamental components like actuators, robotic hands, software, and safety standards are still being rebuilt. This assessment stands in stark contrast to voices claiming that general-purpose foundation models are about to ignite a robotics revolution.

The Boundaries of Physical AI Extend Far Beyond Robots

Vishwanath's core argument is that the scope of Physical AI is far broader than what the word "robot" encompasses. She points out that physical AI focuses on how intelligent systems perceive, reason, and interact with the physical world — including but not limited to autonomous mobile robots, industrial automation arms, drones, and even embodied intelligent agents across scenarios like smart manufacturing, logistics, and healthcare.

In other words, robots are merely a subset of physical AI. Equating physical AI with humanoid robots or robotic arms severely limits our understanding and investment positioning in this technological wave. Vishwanath argues that true breakthroughs will occur when multiple forms of physical systems simultaneously achieve intelligence leaps — not when a single category explodes.

Embodied Intelligence is a key concept for understanding physical AI. It emphasizes that intelligence cannot exist apart from a body — an agent needs to perceive its environment through sensors, exert actions through actuators, and learn and make decisions through real-time interaction with the physical world. This idea originates from embodied cognition theory in cognitive science and was later introduced into the AI field. Unlike traditional pure-software AI (such as language models), embodied intelligent agents must handle continuous control signals, uncertain physical dynamics, and safety constraints. What Vishwanath refers to as "autonomous mobile robots, industrial arms, drones" are all different carriers of embodied intelligence. Understanding this concept helps grasp the investment logic of physical AI: value lies not in any single form factor, but in whether the perception-reasoning-action closed loop can be replicated across diverse physical systems.

The "Lego Era": Fundamental Components Are Still Being Reshaped

She uses the "Lego era" as a metaphor for the current state of robotics technology, meaning the industry has not yet formed stable, reusable standard modules. Actuators, dexterous hands, underlying software frameworks, safety certification standards — every "building block" that constitutes a robot is rapidly iterating or even being redefined.

This judgment directly challenges the optimistic expectation that "general foundation models will soon make robots take off." Vishwanath's logic is: even if the AI brain is smart enough, if the physical body's components lack reliability, standardization, and cost advantages, scaled deployment is out of the question. The current bottleneck is not in algorithms, but in the maturity of hardware and system integration.

According to analyses and interpretations of this viewpoint by Bilibili creators and YouTube channels, many industry insiders agree with this "components-first" assessment, believing that 2026 will still be a phase of laying the foundation for physical AI rather than a harvest period.

Sim-to-Real Transfer is a key technical pathway mentioned by the optimists. The idea is to train robot policies in virtual environments and then deploy the models onto real hardware. The advantage is that simulation can generate massive training data at low cost and in parallel, avoiding the high costs and safety risks of real-machine sampling. However, the physical discrepancies between simulation and reality (such as friction, deformation, sensor noise) lead to a "reality gap" that causes policies to fail on real hardware. Domain randomization, system identification, and adaptive control are common approaches to bridge this gap. The maturity of this technology directly influences the direction of the debate over whether "general foundation models can bypass hardware bottlenecks": if transfer is reliable enough, dependence on high-precision actuators and sensors may decrease; conversely, hardware components will remain an insurmountable bottleneck.

Investment Logic: Bet on Components, Not Complete Machines

As a newly minted investor, Vishwanath's investment perspective clearly leans toward the infrastructure layer. She focuses more on teams that provide "Lego bricks" for physical AI — such as high-performance actuators, tactile sensors, real-time motion control software, and safety certification solutions.

This strategy is similar to the investment logic of the internet era: selling shovels during a gold rush. While complete-machine manufacturers are still struggling to explore form factors, costs, and scenarios, component suppliers may see scaled demand earlier. Vishwanath emphasizes that the supply chain for physical AI is far more complex than traditional software, involving mechatronics, materials science, and functional safety — fields with longer innovation cycles but deeper moats.

Safety Standards: The Underestimated Deciding Factor

In Vishwanath's discussion, safety standards are repeatedly mentioned yet often overlooked by the industry. Once physical AI systems enter human living and working spaces, functional safety, liability attribution, and certification systems become prerequisites for commercialization. Currently, robot safety standards remain fragmented, with vastly different requirements across regions and application scenarios.

She believes that whoever can participate in defining and first comply with these standards will gain an early advantage in the scaling race for physical AI. This viewpoint complements discussions that focus solely on AI model performance, reminding us that intelligence in the physical world must be built on a foundation of trustworthiness and safety.

Functional Safety is a hard threshold for the commercialization of physical AI, with core standards including ISO 13849 (safety-related control systems for machinery) and IEC 61508 (functional safety of electrical/electronic/programmable electronic safety-related systems). For collaborative robots, ISO/TS 15066 further specifies force and pressure limits for human-robot coexistence. These standards require systems to enter a safe state upon failure and to possess verifiable Safety Integrity Levels (SIL). The challenge for physical AI is that learning-based methods (such as neural networks) are essentially black boxes, making it difficult to prove their safety using traditional formal methods. This has spawned new research directions such as "safe learning" and "runtime monitoring." Whoever can first solve the compatibility between the interpretability of AI decisions and safety certification will hold the key to scaling physical AI.

Consensus and Divergence Across Multiple Perspectives

Synthesizing analyses from multiple sources, Vishwanath's "Lego era" judgment has resonated widely, especially among hardware engineers and manufacturing practitioners. The consensus is that physical AI will be deployed more slowly than pure software AI, and that the maturation of fundamental components requires patience.

The divergence lies in the timeline. Some optimists believe that the generalization capability of foundation models is sufficient to bypass some hardware bottlenecks, accelerating robot evolution through sim-to-real transfer. Vishwanath does not completely dismiss this pathway, but she insists that the complexity of physical interaction means the "building blocks" must be assembled steadily, one by one.

Regardless of which viewpoint ultimately prevails, Vishwanath's reminder is worth heeding: physical AI is more than robots — it is a systems engineering endeavor that requires investors, developers, and standards makers to advance together with a broader vision.

One-line Summary

Physical AI far exceeds robots; the industry is still in the 'Lego era,' with investment opportunities in components and safety standards.

Paragraph Summary

Amulya Vishwanath, former head of product marketing for Nvidia's robotics platform and now a physical AI investor, proposes that physical AI encompasses far more than robots — it covers all types of embodied intelligent systems. She believes robotics is currently in a 'Lego era,' where fundamental components like actuators, dexterous hands, software frameworks, and safety standards are still rapidly iterating and have yet to form stable, reusable standard modules. This judgment challenges the optimistic expectation that 'general foundation models will soon ignite a robotics revolution,' pointing out that the bottleneck lies in hardware and system integration rather than algorithms. In terms of investment logic, she favors the infrastructure layer — high-performance actuators, tactile sensors, real-time control software, and safety certification solutions — rather than complete-machine manufacturers. Safety standards are repeatedly emphasized as an underestimated key to commercialization. The industry diverges on the deployment timeline, but component maturity is a consensus bottleneck.