NVIDIA published a technical walkthrough, using its Microduck reference robot, that splits a robotics workflow between AI agents and reinforcement learning. On an RTX Spark workstation, Hermes and Nemotron 3.5 Bolt turn a prompt into a step-by-step project plan; a local image model renders a concept of the obstacle course, and an agent extracts an asset inventory with cropped references. A second agent, Astra, isolates and upscales those crops, models geometry in FreeCAD, applies materials in Blender and exports USD assets for assembly in Isaac Sim, with Newton simulating the robot's motion and balance. Training then runs in Isaac Lab: a pretrained locomotion policy receives additional training on ramps and obstacles using head-camera depth and robot-state observations, across 4,096 parallel environments on an RTX PRO 6000 Blackwell Workstation Edition GPU. The trained policy is evaluated locally in Isaac Sim, recording course completion, falls, collisions and where progress stops. The post says detailed hardware deployment is out of scope, and describes modelling servo command delay, backlash and voltage sag, plus domain randomisation, to narrow the simulation-to-reality gap.
Our takeSeparate what the agents did from what was learned. The agents handled the tedious middle: reading robot documentation, sketching the course, generating meshes and materials. The learning stage stayed conventional reinforcement learning with hand-specified observations, rewards and a single large parallel run. The post reports no success rate, only that completion, falls and collisions were recorded, and the transfer work rests on modelling servo delay, backlash and voltage sag plus domain randomisation. Those are the terms that decide deployment, so the agent layer removes preparation effort,