On 9 October 2026 I attended the NVIDIA Physical AI workshop at Netaji Subhas University of Technology (NSUT) in Delhi. Physical AI is the same loop as software AI with a body attached: a model reads the world, plans, and acts, and the feedback comes from a machine that can fail in the physical world.

This post covers what the day included. Where I do not have a result to report, the post says so.

The day at a glance

The workshop ran from 9:00 a.m. to 5:00 p.m. at NSUT, organised by the Department of Computer Science and coordinated by Dr. Gaurav Singal. The agenda had an overview of the NVIDIA open model ecosystem, a live demo with Addverb, two hands-on labs, and a closing Q&A on proof-of-concept scoping with NVIDIA solutions architects.

The lab lead was Ninad Madhab, Senior Solutions Architect for edge AI, robotics, and industrial AI deployment. The labs ran on NVIDIA Brev cloud GPUs, and Brev credits were provided.

The stack in one picture

The overview covered a pipeline with three stages: simulate, train, deploy.

  • Simulate: Isaac Sim and Isaac Lab build environments; Cosmos generates training video.

  • Train: GR00T post-training on DGX.

  • Deploy: inference on Jetson Thor at the edge, with real-world data fed back into simulation.

  • Simulate. Isaac Sim and Isaac Lab generate environments and training data. Cosmos generates synthetic video: Predict produces future world states, Transfer produces photoreal data from simulation input, and Reason is a physical-reasoning vision-language model.

  • Train. Isaac GR00T N1.7 is an open vision-language-action model of about 3 billion parameters, released in April 2026. It pairs a vision-language backbone with a diffusion transformer action head. Post-training adapts it to a specific robot and task.

  • Deploy. Jetson AGX Thor (T5000) runs inference at the edge. It is a Blackwell GPU with up to 2070 FP4 sparse TFLOPS and 128 GB of memory, with a 40–130 W power envelope.

The slides ended the pipeline with a feedback loop: data from the deployed robot goes back into the simulation and training stages, so deployment feeds the next round of training rather than closing the project.

Slides also described the robot control stack as layers running at different rates: high-level reasoning at 1–5 Hz, perception and planning at about 10 Hz, and real-time control at 100 Hz to 1 kHz.

The Thor and Addverb demo

The showcase used three machines: Jetson Thor at the edge, an RTX PRO 6000 for simulation, and a vision-language model with a vision-language-action policy running hardware-in-the-loop. Addverb, represented by Rajesh Kumar, Head of Advanced Robotics, provided the robotics side.

[Add what the robot did and the response time you measured. If you did not measure it, say so here.]

Lab 1: an end-to-end pipeline on Brev

Lab 1 covered the full path on one cloud GPU: fine-tuning Cosmos, developing a VLM/VLA model, feeding it sensor input from the browser's microphone and camera, and a walkthrough of edge deployment.

[Add what you actually ran in Lab 1 and where it stopped, if it did.]

Lab 2: humanoid simulation with Isaac GR00T

Lab 2 was a Developer Learning Institute course on humanoid simulation. The steps were Isaac Lab-Arena setup, converting HDF5 data to LeRobot format, GR00T 1.7 post-training, and closed-loop pick-and-place evaluation.

[Add whether the pick-and-place evaluation ran, how many tries it took, and the result. If it did not run, say what stopped it.]

What was hard

[Add the exact errors you hit and how you got past them. Leave this section empty rather than guessing.]

What I would build next

[Add your proof-of-concept idea and the feedback from the NVIDIA solutions architect.]

Frequently asked questions

What is Isaac GR00T? Isaac GR00T is NVIDIA's open vision-language-action model family for humanoid and general-purpose robots. N1.7 is about 3 billion parameters, with a vision-language backbone and a diffusion transformer action head.

What is Cosmos used for in physical AI? Cosmos is NVIDIA's family of world foundation models. Predict generates future world states as video, Transfer generates photoreal synthetic data from simulation, and Reason is a physical-reasoning model.

Why run inference on Jetson Thor instead of the cloud? A robot has to react in real time. Running the policy on the device avoids a network round trip in the control loop.

What hardware do I need for the labs? The labs ran on NVIDIA Brev cloud GPUs, so the laptop only needed to run a browser and a shell. The recommended connection was 10 Mbps down and 2 Mbps up. Prerequisites were Python, Docker, Git, and basic 3D concepts.

Is this workshop useful if I do not build robots? The simulation, data, and deployment patterns carry over to any system that acts on sensor input. The robot itself is the part most people will not need.

Thanks

Thanks to NVIDIA, Ninad Madhab, Addverb and Rajesh Kumar, and NSUT and the Department of Computer Science for hosting the day.

Volanea is an email platform for transactional sending, campaigns and automation.