How Close Are We to a ChatGPT Moment for Robots? Nvidia Weighs In | Disrupt 2026

In this talk

1. 0:44

   What physical AI has in common

   Carpus groups vehicles, humanoids, factory robots, and smart buildings under physical AI because they all operate in real time amid unexpected conditions. He describes a data flywheel: use real-world deployment to improve simulation, then use simulation to refine and retrain models. When assessing startups, he prioritizes customer understanding, product-market fit, and teams that can integrate tools and deliver customer value independently.

2. 3:22

   Data, simulation, and deployment

   Carpus says the right investment depends on where a robot fails: gather more real-world data for unfamiliar situations, generate synthetic data for known but difficult cases, or retrain when the data is good but the model is not learning. He expects companies to combine captured and synthetic data, with structured, metadata-rich collections supporting task-specific training. He also sees reusable, sim-ready 3D data as a valuable long-term resource.

3. 7:13

   Specialized robots and world models

   Carpus expects focused robots to reach return on investment sooner, while general-purpose models could eventually perform many specific tasks. He describes world action models as tools for reasoning about scenes, transforming video conditions, and generating possible future sequences for training robot policies. He says current world models may lack physics understanding, though that limitation is being actively addressed.

4. 11:06

   Reliability, form factors, and the ChatGPT moment

   Carpus expects broad returns from humanoids to take three to five years, citing regulatory and liability barriers, while wheeled bimanual robots may be easier to deploy. Legs can help in places such as construction sites with uneven ground, and human activity video offers potential training data for human-shaped robots. He defines a robotics ChatGPT moment as a robot learning a task in a new environment without an engineer guiding it, including recovering from mistakes.

5. 16:10

   What to watch in the robotics ecosystem

   Carpus points to larger context windows and parallel simulation as ways to improve robot learning and policy testing. He cautions startups against putting a third-party foundation model at the center of their value proposition, and says they should own their core model and post-training. He identifies multiphysics simulation, secure collaboration across different robot fleets, and fast feedback from field data as important challenges and opportunities.