The article is dated October 21, 2026 and describes Microduck as an open-source robot from Pollen Robotics and Hugging Face.
The provided Isaac Lab tasks include the robot model, observations, actions, rewards, and domain randomization. The policy uses 61 observation values and produces 14 joint-position outputs. The rough-terrain example runs 4,096 parallel environments; the authors report up to 1.25 million steps per second in a 16-GPU configuration with rough terrain and backlash enabled.
Isaac Lab playback exports TorchScript and ONNX files. Microduck's onboard runtime loads the ONNX policy. Microduck Arena accepts compatible submissions with 61 observations, 14 actions, and a maximum file size of 64 MiB. The guide describes randomizing physical parameters including friction, mass, encoder offsets, battery behavior, latency, terrain, and backlash.
The source does not establish a price for the robot or hardware needed, independent transfer-success rates, training time, or comparative results against other sim-to-real systems.