NEURA Robotics Executive Talk | Nakul Duggal (Qualcomm) on Physical AI

In this talk

1. 0:21

   Why physical AI needs more than vision

   Duggal says that automating human tasks in varied physical settings is a new challenge for AI. Vision-language-action models alone have not handled real industrial tasks well, and relying on video alone requires too much training data. The partners want robots to learn new tasks in under three hours.

2. 9:13

   A brain and a reactive nervous system

   NEURA’s executive describes a two-part approach: a model that understands tasks, plus fast neural networks that respond to touch and force. A robot could use fingertip feedback to grip a cup without dropping it, adjusting as the liquid shifts. These quick reactions help with changing conditions and reduce the need to slow down to compensate for model delays.

3. 6:28

   Skills that can transfer across devices

   The executive argues that touch and force feedback can help skills work across different robot designs, since they need not depend only on precise visual knowledge of each device’s shape. Duggal also discusses using connected cameras and other devices to give robots useful information about their surroundings. The group describes edge processing and local networks as important where cloud delays are too long.

4. 22:17

   Training in realistic settings

   NEURA’s executive describes a three-level architecture: the nervous system, the brain and an extended brain that draws on information about the robot’s environment. NEURA Gyms recreate settings such as kitchens or industrial workspaces to train and evaluate robots. The executive says the goal is to make training data-efficient, with some tasks taking less than two hours, and accessible without specialist expertise.

5. 26:49

   Building an open developer ecosystem

   Duggal points to Qualcomm’s work with Arduino and its community of more than 30 million developers as a way to put capable tools in more hands. The executive describes NEURAVERSE as a platform where companies can connect sensors to AI models, train and deploy them in secure environments, and keep private know-how to themselves. Both argue that progress depends on organizations building together rather than one company owning every solution.