Interview at the NVIDIA GTC with Spencer Huang: Humanoid Robotics

In this talk

1. 0:19

   Humanoids before full autonomy

   Huang says humanoid autonomy is still developing, so teleoperation offers practical uses today. He points to remote data center checks, where an engineer could respond to an alert from home, and to perimeter security. Operating robots also helps teams improve their bodies and mechatronics while autonomy advances.

2. 2:16

   Simulation depends on the task

   Huang expects computer vision in factories and smart buildings to grow before robots take on more complex physical interaction. He says simulation already works well for virtual commissioning of conventional robots, but dynamic control and contact with objects leave harder sim-to-real problems. The host notes that accurate physics, including modeling cloth, matters for closing those gaps.

3. 5:04

   Agents could speed up robot development

   Huang proposes using agents to generate and test simulation scenarios, vary objects, materials, and lighting, and search for edge cases in trained policies. He also describes automating data review: agents could identify unusual driving data, augment it, and help retrain models. He sees this as a way to run more experiments in the same amount of time.

4. 6:48

   Robots and natural language over the next five years

   Huang expects smarter household devices and mobile manipulators to appear within one to three years, with robots capable of real tasks developing over roughly five years. He argues that agentic systems and natural language could make robots easier to task and program, reducing reliance on specialist programming that can take months. He also says people should set the principles for functional safety, with AI helping test whether robots follow them.

5. 9:38

   AI as assistance and a new kind of work

   Huang expects some retraining as AI spreads, but frames it as creating new roles and helping people do more. He compares AI-assisted radiology to expanded capacity for treating patients, rather than simply eliminating radiologist jobs. He says natural language could make learning new work easier, while agents help teams manage more tasks and accelerate robotics research.