Core Viewpoint - The research teams from Peking University and Renmin University have made significant breakthroughs in the field of general humanoid robot motion generation, introducing the innovative data-model collaborative scaling framework, Being-M0 [1][2]. Group 1: Motion Generation Dataset - The team has created the industry's first motion generation dataset, MotionLib, with over one million action sequences, significantly enhancing data acquisition efficiency through an automated processing pipeline [4][7]. - MotionLib includes over 1 million high-quality action sequences, achieving a scale 15 times larger than the current largest public dataset, thus overcoming the scale bottleneck in motion generation [10]. Group 2: Large-Scale Motion Generation Model - The proposed large-scale motion generation model demonstrates significant scaling effects, validating the feasibility of the "big data + big model" approach in human motion generation [5][13]. - Experiments show a strong positive correlation between model capacity and generation quality, with a 13B parameter model outperforming a 700M parameter model in key metrics [13][14]. Group 3: Motion Redirection Across Platforms - The Being-M0 team has innovatively integrated optimization and learning methods to efficiently transfer motion data to various humanoid robots, enhancing cross-platform adaptability [6][20]. - A two-phase solution is proposed for cross-modal motion transfer, ensuring high-quality generated data while maintaining real-time performance [21]. Group 4: Future Directions - The Being-M0 project aims to continuously iterate on humanoid robot capabilities, focusing on embodied intelligence, dexterous manipulation, and full-body motion control, ultimately enhancing the general capabilities and autonomy of robots [22].