From Robot Development to Deployment with Isaac GR00T & Jetson Thor

In this talk

1. 4:42

   GR00T and the humanoid development stack

   Brandon Neal of NVIDIA describes GR00T 1.7 as an open vision-language-action model, pretrained on roughly 32,000 hours of human demonstration data. He highlights its Cosmos vision-language backbone and production-oriented license, then outlines NVIDIA’s development platform for data collection, training, evaluation, and deployment on Jetson Thor.

2. 7:15

   Training MOI for industrial work

   Wei says Noble Machines builds MOI as a full-stack robot for hazardous and physically demanding industrial jobs. The team trains whole-body control in Isaac Lab and uses GR00T for autonomy, with a pipeline for collecting hardware data and fine-tuning policies. In demonstrations, MOI handles a sustained 50-pound payload and adapts as the load and its distribution change.

3. 20:59

   Closing the sim-to-real gap

   Wei says heavy industrial robots make sim-to-real transfer especially challenging, including when simulated actuator output exceeds what the real hardware can deliver. Noble Machines uses robot data to improve simulation fidelity and applies physics knowledge to reward design and curriculum, randomizing factors such as center-of-gravity distribution. The company is also working with ADLINK on a Thor-based industrial computer designed for demanding environments.

4. 30:03

   Seeed’s end-to-end learning path

   Elaine and her Seeed Studio colleagues present a roughly 20-to-30-hour course using the ReBot arm to teach teleoperation, data collection, simulation, training, validation, and deployment. Their workflow combines real robot demonstrations, Isaac Sim synthetic data, and Cosmos-generated video augmentation to train GR00T 1.7 policies. The course is free; learners can buy the hardware package, and simulation assets are also available in Isaac Sim.

5. 35:08

   From demonstrations to real robot policies

   Seeed demonstrates a leader arm controlling a follower arm while Isaac Sim mirrors the motion, then shows the robot performing a table-cleaning task. The team says the full course requires physical hardware, though the simulation scene can be explored without it; the pipeline also supports other policies through LeRobot. Johnny adds that seven XR demonstrations produced roughly 60 to 70 percent task success in one example, and highlights TensorRT and Jetson Thor for deployment latency.