Kartik Sachdev

From Data to Deployment: A Founder's Blueprint for Physical AI

The core message of the provided transcript is that physical AI—bringing AI from screens into robots, machines, vehicles, and industrial systems—represents a much larger and harder deployment frontier than current LLM-based information technology. After the moderator introduced the masterclass format and the two NVIDIA speakers, Olga framed the opportunity: LLMs such as ChatGPT already support tasks like email, search, coding, business communication, and presentation building, but they still cannot perform physical tasks such as laundry. Robotics bodies, in her framing, become the platform that allows AI to enter the physical world. She contrasted roughly $5 trillion in information-technology value with around $100 trillion across sectors such as agriculture, automotive, construction, energy, healthcare, life sciences, logistics, and manufacturing if physical AI reaches those domains. Olga emphasized that physical AI is not as simple as impressive robot demos can make it appear. Like other AI systems, it depends on large amounts of data, but collecting and labeling robotics data is especially difficult. She compared older computer-vision workflows—manually drawing bounding boxes or labeling cats and dogs—with the much harder task of teaching robots physical skills. The session introduced teleoperation and imitation learning, where humans demonstrate tasks such as pick-and-place, screwing, cable work, or logistics separation, and reinforcement learning as the two main training paradigms currently used in physical AI. The speakers also noted that startups usually do not begin from scratch; they often fine-tune foundation models, while founders building a robot “brain” may need to develop their own model. The session then moved into NVIDIA’s “three computers” approach as a practical blueprint from data to deployment. The first computer is large-scale training infrastructure, such as DGX-class data-center GPUs, used to train models like Cosmos. The second computer supports fine-tuning, validation, and simulation through tools such as Isaac Sim and Isaac Lab, with vertical-specific building blocks for autonomous vehicles, surgical robotics, medical devices, manufacturing, visual agents, humanoids, and manipulators. The third computer is deployment hardware at the edge, from Jetson Nano to Thor, where models must be optimized to run efficiently and cost-effectively on robots, edge devices, or factory floors. Kartik then deepened the discussion around data regimes for physical AI. He explained that robotics does not have a single internet-scale data source equivalent to what LLMs used; instead, physical AI needs a spectrum of datasets, including teleoperated robot trajectories, videos of humans performing tasks, simulation data, real-world autonomous-vehicle data, and synthetic datasets. He highlighted that synthetic data is valuable because real-world collection cannot cover enough rare or expensive edge cases, such as unusual weather or hard-to-recreate scenarios. The provided transcript ends while he is describing open-source USD simulation-ready assets on Hugging Face, including factory-style assets and community usage.