Amazon, NVIDIA, and a new "physical AI" fellowship

In this talk

1. 0:58

   What physical AI means

   Rasheed describes physical AI as combining spatial awareness with AI so systems can act in real environments. She says language models can contribute, but their reliability depends partly on training data quality, and safety must be built into systems that operate near people.

2. 4:59

   Why the field is gaining ground

   Rasheed points to advances in machine learning and connected devices, workforce shortages, and an aging population as reasons physical AI is drawing attention. She estimates that more than 2.5 billion people do physical work, accounting for nearly $50 trillion in yearly output. Lower component costs, improved sensors, and robotics processors are also helping make development more practical.

3. 7:56

   Data and precise movement remain hard

   Training requires rich, industry-specific spatial data, which may mean adding instruments to the environment to collect it. Robots also need adaptable control: picking up a berry calls for different handling than picking up an orange, while surgical work demands especially precise movement and safety.

4. 10:12

   Fellowship startups tackle varied work

   The first cohort includes eight startups with significant computing needs. Rasheed highlights Bedrock Robotics, which equips existing construction machinery for autonomous operation; the host also describes Diligent Robotics, whose socially capable robot is used in 25 hospitals and has expanded into senior living.

5. 12:53

   General and specialized robots, with early results

   Rasheed expects both broadly capable robots and systems tailored to particular tasks, such as healthcare or specialized manufacturing. She says Amazon fulfillment operations have seen 25% efficiency gains from automation, and the fellowship may include enterprise companies in 2026.