AI accelerates robotics development: virtual training speeds up complex tasks

A developer demonstrated the rapid training of a virtual robot, Microduck, using OpenAI’s GPT-6 Sol to identify Bengal cats. This simulation, which took mere hours, showcases how advanced models can translate 3D world understanding into actionable code and agentic robot training, a process that previously required weeks. This rapid prototyping mirrors broader advancements in the field. At the Conference on Robot Learning (CoRL) in Austin, Texas, Nvidia researcher Jim Fan noted that AI control mirrors software agent functionality. Professor Ken Goldberg and his students at UC Berkeley utilized OpenAI’s Astra to generate a full 3D simulation from a video clip in under an hour, demonstrating Astra's simulation capabilities. Furthermore, the Aspire agentic framework, recently open-sourced by Fan, Goldberg, and others, allows robots to discover new skills, such as Astra solving a Rubik’s cube. Goldberg also showed Astra controlling a robot arm to paint the Golden Gate Bridge. Despite these leaps, the author noted that real-world deployment remains challenging, though the speed of AI training suggests a significant future impact on manufacturing and e-commerce.