XNews

Isaac Robot Learns from Videos, Cuts Teaching Time 210x, Open-Source

Isaac, a new robot learning system, reduces the human teleoperation teaching time from 5,900 hours to just 28 hours by first watching 1 million hours of online video, then fine-tuning with human correction—a 210x efficiency gain. It uses a Mixture of Experts (MoE) architecture that activates only 1–2 experts for simple scenes, cutting compute cost 8.5x versus comparable models. Isaac handles visual question answering, object counting and tracking, and directly controls over 35 robot arms and form factors. With a single demonstration of a new task like chess, its adaptation and error correction outperform competitors by 7–10 times. The model weights, training methodology, and inference code are fully open-source.

Source posts