Dr. Ashutosh Saxena, Founder and Chief AI Officer of TorqueAGI, takes the stage at NVIDIA GTC 2026 to present how TorqueAGI and NVIDIA technology work together to bring physical AI to enterprise-grade robotics.

Making ice cream is easy for humans. For robots, it exposes a hard problem. Long-horizon tasks require memory, planning, and coordination across many steps, not just perception and control.

Cross-embodiment learning shifts robotics from memorizing robot-specific datasets to learning transferable physical abstractions. The result: orders of magnitude less data and far faster deployment across different machines.

This video shows TorqueAGI’s edge native foundation model running fully on device, operating in real time with no data sent off platform. The model runs live, observing the environment, reasoning over it, and triggering behavior under real world constraints

Cross-embodiment learning lets robots build on each other’s experience instead of starting over every time, unlocking their true potential. No retraining. No custom scripting. This was the idea behind Robobrain and the foundation that TorqueAGI is built on.

At TorqueAGI, we are building the Physical AI foundation model that enables this structured decision-making. It lets humanoids handle complex workflows in warehouses, manufacturing, and logistics in the real world today.

VLAs can generalize in the cloud. Robots need real-time on the edge. The only architecture that works puts the entire intelligence pipeline on the GPU. Three steps make this possible.