Gasgoo Munich-On April 28 (US Pacific Time), MagicLab unveiled its proprietary world model, Magic-Mix, at the Global Embodied AI Innovation Conference.
According to the company, the Magic-Mix world model is built around two core engines. The Magic-Mix WAM handles physical environment understanding, spatial reasoning, and action decision-making, while the Magic-Mix Creator serves as an offline data generation engine. It produces large batches of training samples to continuously drive model training and capability iteration.
Image Credit: MagicLab
Together, these modules create a closed loop of "massive data generation, model training, result feedback, and data regeneration." As a result, Magic-Mix is not a static model but a dynamic system. It enables robots to learn continuously in both real-world and simulated environments, constantly correcting themselves to gradually improve their adaptability to complex tasks.
According to MagicLab President Gu Shitao, the company collects roughly 16,000 data points daily, with its high-quality data scale exceeding 1 million hours. Through continuous data synthesis, it has achieved a 10,000-fold expansion in data volume. The core value of Magic-Mix Creator lies in using mass synthetic data to reduce reliance on real-world robot collection, providing a steady stream of high-quality datasets for large model training.
