Source time Edition date
I Saw the Future of Robotics in a Virtual Duck That Chases My Cat
我在一只追我的猫的虚拟鸭子身上看到了机器人技术的未来
OpenAI's model lets the Microduck learn to waddle and play cat butler, but buyers must weigh playful home robotics against how much autonomy they will grant a device that shares their living space.
Judgment
Industrial automation firms and home-robot buyers must decide whether to back agentic simulation frameworks such as Aspire now, trading proven reliability for a training speedup that took a virtual robot from weeks to hours.
Basis
Ken Goldberg at UC Berkeley says Astra produced a full 3D simulation with camera, lighting, background and materials “within an hour”; his lab gave it access to Nvidia’s Newton tool. Will Knight trained a virtual Microduck with OpenAI’s GPT-6 Sol, Nous Research’s Hermes and an Nvidia DGX Spark in “just a few hours,” versus weeks previously, using Nvidia’s Isaac Sim to watch videos and suggest training tweaks. Jim Fan says LLMs can call visual language action models much as software agents call programs, and this week Fan, Goldberg and others open-sourced Aspire; GPT-6 Astra solved a Rubik’s cube after reasoning over failed twists. The limit is reliability: the duck’s vision system “does not work well enough to be deployed in a real product,” and robots still require significant engineering expertise.
Unclear The source does not say how Aspire performs outside demos, what it costs, or when a physical Microduck will ship.
Next Watch whether the open-source Aspire repository or any industrial robotics firm cited by Goldberg publishes independent real-world reliability results.