从“炫技”转向“干活”，轮子比双足更吃香......高盛总结了WAIC人形机器人最新趋势

Core Insights - The 2025 World Artificial Intelligence Conference (WAIC) signals a shift towards practical commercialization in the humanoid robotics sector, moving away from mere technological showcases to robots that can perform real tasks [1][2] - The trend of adopting wheeled chassis robots, similar to Automated Guided Vehicles (AGVs), is becoming mainstream, indicating a focus on immediate commercial viability rather than complex bipedal designs [2][3] - The number of showcased robotic products has significantly increased, with over 60 products presented compared to 25 static prototypes last year, demonstrating substantial progress in product development [2][4] Industry Trends - The WAIC has expanded in scale, with a 35% increase in venue size to 70,000 square meters and a 60% rise in the number of exhibitors to 800, reflecting enhanced industry investment and government support [2] - The design of working prototypes is increasingly converging towards wheeled solutions, which offer advantages in stability, cost, and energy consumption, making them easier to deploy in flat environments like factories and warehouses [2][3] Application Scenarios - Robots are now being designed for specific applications rather than being generic solutions, with a focus on addressing particular industry challenges [4] - Various sectors are seeing the introduction of robots, including manufacturing, logistics, retail, and healthcare, with specific examples of robots designed for tasks such as power inspection and patient interaction [6] Technical Challenges - Despite advancements in autonomous navigation and dynamic movement, the precision of robotic manipulation remains a significant challenge, with task success rates and operational speeds still lagging behind human capabilities [5] - The integration of visual-language-action (VLA) models and reinforcement learning (RL) is seen as crucial for enhancing the general capabilities and task performance of robots [5] Cost and Data Considerations - The cost of entry-level robots has decreased, with a new model priced at 40,000 RMB, but the overall cost reduction in the industry is not yet significant, with full-sized robots still priced between 400,000 to 500,000 RMB [6] - High-quality real-world interaction data is essential for training VLA models, but collecting such data is expensive, leading companies to adopt mixed strategies that combine real and synthetic data for model training [6]