Ex-Nvidia Robotics Exec Amulya Vishwanath Says Physical AI Is Stuck in Its 'Lego Era'

Physical AI is being held back by missing fundamentals, not by a shortage of ambitious demos, according to Amulya Vishwanath, who led product marketing for Nvidia's robotics platforms for seven years before moving to the investor side of the industry. In her view, robotics today is in a "Lego era": actuators, robotic hands, software stacks and safety standards are all still being rebuilt from the ground up. That assessment cuts against the widely repeated expectation that a single general-purpose foundation model will soon push robots into mainstream deployment.

The Lego comparison matters because it reframes where the bottlenecks actually sit. A foundation model can plan and reason, but it cannot compensate for a gripper that lacks the dexterity to pick up an object or a safety framework that no regulator has signed off on. Until those components are standardized and interchangeable, Vishwanath argues, each new robot deployment remains a bespoke engineering project rather than a product that can scale. Her time inside Nvidia gave her a close view of how fragmented the supply chain remains beneath the polished launch videos.

That fragmentation is also why she pushes back on defining physical AI as a humanoid race. Robots are one expression of the category, but the same perception, control and simulation techniques apply to warehouse automation, agricultural equipment, surgical tools and industrial inspection systems. Investors who narrow their thesis to bipedal machines risk missing the less photogenic segments where deployment is already generating revenue. Vishwanath's position echoes a broader shift among robotics funders toward component makers and middleware rather than full robot builders.

Nvidia itself has spent the past several years positioning around this layer of the stack, selling simulation tools, training hardware and development platforms to robot makers rather than building the robots. That strategy has made the company a central supplier to nearly every physical AI startup, which in turn gives former employees like Vishwanath an unusually broad map of who is building what. Her move from that vantage point into investing suggests she expects the returns to accrue to the picks-and-shovels side first.

The practical implication for teams working in the space is that the near-term roadmap should assume non-standard parts and immature safety rules, not a plug-and-play foundation model. Companies that invest in reliable low-level components and simulation pipelines may find themselves better positioned than those chasing end-to-end autonomy. The open question is how quickly the Lego pieces actually snap together — and whether the first general robot foundation model arrives before or after the hardware beneath it is ready to carry it.