nvidia open-sourced a full pipeline for on-device robot control – no data center in the loop

our ai host mira broke down @NVIDIARobotics' new technical walkthrough – cosmos 3 edge is a 4b world model trained for robot manipulation that runs entirely on the robot's own computer, jetson thor, without needing a remote gpu.

the model fits in about 9 gb and predicts the robot's next moves in roughly 1.5 seconds – fast enough for the arm to keep moving continuously without waiting. after each cycle, the robot corrects its plan based on where it actually is, not where the previous prediction expected it to be.

it learned from 76k human-controlled robot demonstrations spanning 350 hours across 86 different tasks. the entire pipeline is open – model weights, training data, code, and a simulation environment for testing before you run anything on a real robot.

success rate in simulation: 22.9% versus 36.8% for the larger cosmos 3 nano – but nano doesn't fit on the robot. that's the trade: real-time autonomy on the machine at the cost of some accuracy.

this isn't just a model release – it's a full reproducible workflow from data to deployment, and it supports multiple robot platforms. nvidia open-sourced the entire blueprint.