Workflow
近亿元融资背后:灵宝CASBOT的技术务实主义与垂直场景突围
机器人大讲堂·2025-06-27 02:20

Core Viewpoint - The article discusses the recent financing of humanoid robot company Lingbao CASBOT, highlighting its significance in the competitive humanoid robot industry and its implications for technological evolution and commercialization [1]. Financing and Investment - Lingbao CASBOT completed a new financing round amounting to nearly 100 million RMB, led by Lens Technology, with participation from Tianjin Jiayi and existing shareholders [1]. VLA Model Engineering Challenges - The Vision-Language-Action (VLA) model faces challenges in transitioning from theoretical research to engineering applications, particularly in maintaining generalization while meeting specific task accuracy and stability requirements [2]. - CASBOT collaborated with the Chinese Academy of Sciences to propose the ConRFT method, which enhances the VLA model's task execution capabilities through reinforcement learning [2]. ConRFT Methodology - The ConRFT method employs a two-stage optimization strategy, including an offline fine-tuning phase that combines calibrated Q-learning and behavior cloning to train the Q function with limited data [2][3]. - The online fine-tuning phase incorporates a human-in-the-loop mechanism, allowing human operators to intervene and correct the robot's actions, thus improving the model's performance [3]. Performance Metrics - The ConRFT method achieved a 96.3% average success rate across eight real-world tasks, a 144% improvement over traditional supervised fine-tuning methods, and reduced the average task completion steps from 56.3 to 30.7, enhancing efficiency by approximately 1.9 times [5]. Data Collection Strategy - CASBOT established a "three-domain" data collection system, including real machine domain, simulation domain, and human domain, to improve data quality and training efficiency [6]. - The human domain data collection innovatively captures natural human actions, significantly reducing the time required for remote operation tasks [6]. Technical Architecture - CASBOT adopted a layered end-to-end model architecture, allowing flexible resource allocation based on different scenario requirements, enhancing response speed and reducing network dependency [7][10]. - The design reflects practical considerations for deployment environments, ensuring data security and stability in industrial and mining applications [10]. Product Parameter Optimization - CASBOT 02 features specific parameter adjustments, such as height reduction from 179cm to 163cm and computing power adjustment from 550T to 275TOPS, reflecting a strategic shift towards practical application feedback [11][13]. Commercial Strategy - CASBOT focuses on industrial manufacturing and mining energy as primary commercialization directions, driven by market demand and the company's capabilities [15][16]. - The choice of industrial manufacturing is based on rigid demand, quantifiable ROI, and technical adaptability, with applications in quality inspection and shoe manufacturing [16]. Mining Energy Sector - The mining energy sector presents a unique commercialization opportunity, driven by regulatory requirements for robot replacements in critical positions by 2026 [17]. Collaboration and Market Strategy - CASBOT employs a joint R&D model with partners, providing standardized robot products while partners handle specialized modifications and certifications [19]. - The company adheres to a principle of small-scale mass production and delivery, ensuring that products meet clear commercial needs [20]. Industry Trends - The humanoid robot industry is evolving towards mixed technology solutions, with a focus on practical applications rather than singular technological paths [23]. - Innovations in data strategies and product design reflect a shift towards practicality and market maturity, emphasizing effective problem-solving and reasonable ROI [24]. Future Outlook - The humanoid robot industry is still in its early stages, with rapid technological and business model evolution, moving towards more practical and application-focused developments [25].