Jensen Huang: Everything That Moves Can Have FSD, but Robots' Hardest Problem Is Not Motion
Jensen Huang has said that everything that moves will eventually gain autonomy. Not just cars and trucks, but robots, construction machinery, and drones will evolve into intelligent agents that perceive, decide, and act. The comment is not only about the automotive industry; it points to a shared direction for intelligence in the physical world.
The claim borrows Tesla's FSD naming but extends far beyond cars. A car must sense its environment, make decisions, and control motion; an excavator, a robot, or a drone needs the same set of capabilities. If that stack can migrate across categories, it becomes infrastructure for the physical world. This also explains why Nvidia treats autonomous driving and robotics as closely related computing problems.
Nvidia's layout points to one technology stack. Huang places autonomous driving and physical AI in the same narrative, reflecting Nvidia's bet on the next computing platform. Its physical AI efforts cover simulation to deployment: Isaac Sim and Isaac Lab for robot simulation and training, Cosmos world foundation model for physical AI reasoning and action simulation, GR00T as a vision-language-action model for humanoids, and Jetson Thor for edge deployment. Together these tools form a pipeline from synthetic data generation to real-world deployment.
Huang has said autonomous driving taught Nvidia much about helping other domains build robotic systems. By building the entire stack itself, Nvidia learned what kind of chips robotic systems need. That experience now directly informs Nvidia's robotics strategy.
Talent migration from autonomous driving to embodied AI suggests deep overlap. Zhang Tao, CEO of Guangxiang Technology, argues that if a vehicle is a mobile robot, autonomous driving and embodied AI are essentially the same thing. About 50% of autonomous driving infrastructure can transfer to embodied AI, he estimates. The shift is visible in hiring, startups, and research labs moving from AVs to embodied intelligence.
The whole development pipeline—data collection, cleaning, end-to-end model training, simulation validation, and real-machine testing—can be reused. But the transfer is not seamless. Cars need only move and steer on a single plane; humanoid robots have dozens of degrees of freedom. Those differences mean the reusable half must be paired with new work on manipulation, touch, and unstructured environments.
Autonomous driving operates on structured roads, interacting mainly with vehicles, pedestrians, and traffic facilities. Embodied AI must handle irregular, discrete scenarios in industrial lines and daily life. The deeper difference is interaction: for AVs, contact means collision; for robots, grasping and assembly are core tasks. In factories and homes, every object and surface can vary, so robots cannot rely on lane markings or traffic rules.
Manipulation is the hurdle. Wang Qian, founder of Zibianliang, is more direct: autonomous driving models cannot transfer directly to robots because AV difficulty centers on navigation and motion, while robots' biggest challenge is complex manipulation. General robot capabilities fall into motion, navigation, interaction, and manipulation; manipulation is key to entering factories and homes. Locomotion can be improved through mechanical design and control algorithms; manipulation requires the robot to understand and change the physical world.
Running, jumping, and fighting can iterate quickly through body structure, motor power, and balance control. Hand manipulation involves tactile feedback, force-control precision, and multi-finger coordination; each engineering step is harder than locomotion. That is why dexterous hands remain expensive and unreliable.
Data reveals the gap. AV scenarios can accumulate miles through fleet scale, but robot manipulation data must be collected for specific tasks. In logistics sorting, Figure found that eight hours of carefully curated demonstration data can produce flexible policies; quality matters more than quantity. For contact-rich tasks, simulation cannot yet replace real demonstrations.
Simulation is already useful for navigation and locomotion, but physical fidelity remains insufficient for contact-rich dexterous manipulation. Huang's claim that everything that moves can have FSD holds at the navigation and motion level. Perception, planning, and control from AVs can indeed transfer to robots, construction machinery, and drones. But the hardest part of physical AI—dexterous, adaptive manipulation—still lacks the data and models needed for generality.
But for robots to truly work in factories and homes, they need not just to move but to manipulate. Until that hurdle is crossed, physical AI's ChatGPT moment is still half missing. The next breakthrough must come from manipulation, not motion.