Workflow
合成预训练 + 真机对齐
icon
Search documents
2025深蓝智库|银河通用:成为现实生产力
Bei Jing Shang Bao· 2025-05-25 07:59
Core Insights - Galaxy General, founded in May 2023, has rapidly emerged in the humanoid robot sector, showcasing its Galbot G1 robot with a task success rate of 99.97% during the 2025 Zhongguancun Forum [1][9] - The founder, Wang He, emphasizes the importance of synthetic data for training humanoid robots, arguing that relying solely on real data is costly and time-consuming [4][16] - The company aims to drive the practical application of embodied intelligent robots across various scenarios, focusing on real-world utility [10][15] Data Strategy - Galaxy General has established a self-developed synthetic data generation pipeline, allowing for the mass production of diverse synthetic data at low marginal costs, which constitutes over 99% of the training data [5][6] - The company employs a strategy of "synthetic pre-training + real data alignment," which is deemed cost-effective and sustainable even as humanoid robot shipments increase [5][16] - The GraspVLA model, developed in collaboration with academic institutions, showcases the effectiveness of synthetic data, achieving a record training volume of one billion frames of "visual-language-action" data [6][7] Application and Development - The Galbot G1 robot is designed for versatility, capable of performing various tasks in multiple environments, including retail and healthcare [11][12] - The company has initiated strategic partnerships, such as with Suzhou, to implement humanoid robots in commercial retail and advanced manufacturing [10][14] - The "smart pharmacy" solution, which utilizes Galbot G1 for inventory management, is set to be deployed in approximately 100 stores in major cities by 2025 [13] Future Outlook - Galaxy General believes that the development of humanoid robots will continue to face challenges related to data acquisition and hardware capabilities, with a focus on achieving practical productivity [16][17] - The company anticipates significant breakthroughs in humanoid robot capabilities by 2025-2026, potentially leading to widespread adoption in various sectors [17]