NVIDIA’s October 23, 2026, technical walkthrough places RTX Spark at the local development and policy-evaluation stages of its Microduck workflow. Isaac Lab handled policy training using cloud resources, while a separate GPU powered a specified ramp-and-obstacle adaptation.
RTX Spark handled local planning and evaluation
NVIDIA describes Hermes and NVIDIA Nemotron 3.5 Bolt turning a prompt into a project plan on RTX Spark. Astra helps prepare assets and the simulation; FreeCAD and Blender are used to create assets, and Isaac Sim assembles and runs the simulated environment.
NVIDIA’s workflow then uses Isaac Lab for reinforcement-learning training. After training, a selected policy is evaluated in Isaac Sim on local RTX Spark. A policy is a learned set of control decisions that guides the robot’s movement.
NVIDIA’s demonstration traces the workflow from asset creation and simulation through cloud training, local evaluation and physical deployment. NeoTeo has also covered RTX Spark in a separate laptop context in its report on one system’s local AI use.
NVIDIA’s obstacle adaptation used 4,096 environments
For a ramp-and-obstacle policy adaptation, NVIDIA reports using 4,096 parallel environments on an NVIDIA RTX PRO 6000 Blackwell Workstation Edition GPU. Isaac Lab trains a policy through repeated simulated experience; Isaac Sim evaluates a selected policy for course completion, falls, collisions and stalled progress.
NVIDIA’s walkthrough discusses calibration, actuator response, domain randomization and later hardware testing as considerations for carrying simulation results over to a physical robot.
Pollen says owners can retrain on their own machines
Microduck is Pollen Robotics’ 25 cm bipedal robot. Pollen lists 15 motors, a camera, LiDAR, two IMUs and a 50 Hz onboard policy loop. The company says owners can retrain policies on their own machine or through Hugging Face Jobs, then deploy them to the robot. Pollen describes its software stack as licensed under Apache-2.0.
That gives owners options beyond the infrastructure in NVIDIA’s showcase: Pollen describes both local retraining and hosted jobs, rather than requiring RTX Spark for every policy update.
Huang introduced four Microducks at GTC Berlin
Jensen Huang presented four Microduck robots at NVIDIA’s GTC Berlin keynote on October 21, 2026, and gave them commands to sit, stand and sleep. NVIDIA published its technical walkthrough on October 23, detailing the division between local development and evaluation on RTX Spark and policy training in Isaac Lab.
