Tech Talk: An expert conversation on robotics and automation with Spencer Huang, NVIDIA

In this talk

1. 1:37

   What makes this a turning point for robotics

   Spencer says robotics may be approaching a shift like the one that brought AI to everyday users, as more computing power makes advanced models practical. But manipulation lacks the broad training data available for language and images. He sees progress in making robotic capabilities accessible through applications that let users specify tasks within defined limits.

2. 4:55

   Building useful training data

   Robots need to perceive, plan, and act, and each operating domain requires relevant data. Spencer recommends using human demonstrations to capture varied skills, then expanding datasets with synthetic examples and augmentation. The hard part is judging whether synthetic data will actually improve a model.

3. 8:23

   Where industrial robots deliver value

   Spencer points to defect detection as a current use for synthetic data, especially when manufacturers handle many products and rarely seen faults. He also describes mobile robots moving heavy loads and robotic arms handling pick-and-place work. For unfamiliar products, robots must generalize how to grasp items and place them in the required orientation.

4. 10:55

   Adoption barriers and realistic expectations

   Many tasks left to automate require complex, precise sequences that are difficult to program as general skills. Spencer says teams need to combine learning from demonstrations with reward-based training, use simulation to train faster, and address the gap between simulated and real conditions. He cautions that human-level humanoid abilities are not a near-term expectation and that synthetic data needs careful engineering.

5. 15:57

   Preparing for the next decade

   Spencer hopes future robots will learn new tasks in their work environment from demonstrations, instructions, or manuals, then share improvements across a fleet. For industrial teams today, he advises capturing useful data early, developing simulation and 3D expertise, and considering useful robot components such as arms or mobile bases instead of focusing only on humanoids.