NVIDIA details AI-agent workflow for training Microduck in simulation

NVIDIA outlined a workflow for teaching a reference robot called Microduck to walk and navigate obstacles in simulation, tying together AI-assisted planning, asset creation, physics modeling, policy training and local validation. The company said the process is designed to narrow the gap between simulation and real hardware, though final physical deployment remains outside the scope of the post published on 2026-10-23.

Evidence and context

In the developer blog post, NVIDIA said the workflow starts on RTX Spark, where Hermes and NVIDIA Nemotron 3.5 Bolt turn a task prompt into a structured plan for the obstacle course, while Astra helps create and verify assets. The company described the plan as covering research, concept generation, asset inventory, handoff, asset creation and simulation testing. The post said the robot must do more than follow a walking policy: it must perceive its surroundings, keep balance and adjust when obstacles block its path.

The workflow then moves to simulation asset production. NVIDIA said Astra can enhance reference images, build editable geometry in FreeCAD, apply materials in Blender and export USD assets for use in Isaac Sim. The post also pointed to SimReady Central Library, described there as a free library of more than 2,000 simulation-ready OpenUSD assets, as a source of reusable environment components. NVIDIA said the assembled environment combines these assets with Isaac Sim, Omniverse rendering and Newton physics to model robot motion, sensor observations and balance.

For policy training, the company said Isaac Lab runs many simulations in parallel and uses reinforcement learning to improve a shared policy. In the example described, Microduck was trained to traverse a curved ramp and uneven surfaces using head-camera depth and robot-state observations. NVIDIA said the adaptation run used 4,096 parallel environments on an NVIDIA RTX PRO 6000 Blackwell Workstation Edition GPU. The post did not provide comparative performance metrics or indicate whether the resulting policy was tested on physical hardware in the same workflow.

After training, NVIDIA said the policy is validated on RTX Spark in Isaac Sim, where the robot receives simulated observations, predicts actions and is checked for obstacle detection, balance and course completion. The company said evaluation should record falls, collisions and points where progress stops, with rollout videos used to guide further iteration. The post also discussed simulation-to-real preparation, including Better Actuator Models to reflect command delays, backlash and voltage drops, as well as domain randomization to expose the policy to variation in physics and sensor noise.

The article frames the approach as a staged pipeline for robotics development rather than a claim of universal sim-to-real transfer. NVIDIA said further refinement is still needed before running a trained policy on the robot’s onboard computer, and that hardware deployment details were outside the scope of the post.