SIASUN(SZ:300024)2025-11-24 01:46Summary of Key Points from the Conference Call Industry Overview - The conference call discusses the embodied intelligence sector, focusing on the evolution of robotics and AI technologies, particularly the shift from model-driven to data-driven approaches in robot algorithms [1][2][3]. Core Insights and Arguments - Algorithmic Changes: The robotics industry is experiencing a significant transition from model-driven algorithms to data-driven approaches, driven by advancements in generative AI since 2022. This shift allows robots to not only perform actions but also understand and reason about tasks [2][3]. - Main Algorithm Architectures: Three primary algorithm architectures are identified: 1. Hierarchical Control Framework: Established since 1985, separating perception and motion control, still widely used due to its minimal disruption to existing systems [4]. 2. VLA (Vision-Language-Action) Model: Gaining traction among startups since 2023, suitable for interactive scenarios but may need to work alongside hierarchical frameworks in industrial settings for safety [4]. 3. World Model: Focuses on autonomous understanding of the physical world through continuous data, requiring high-fidelity simulations, but faces challenges in practical deployment [4][8]. - Data Acquisition Methods: The industry relies on three main data acquisition methods: 1. Real Machine Acquisition: High-value but costly, involving remote operations and large-scale training environments. 2. Video Learning: More cost-effective, using real video recordings to train robots. 3. Simulation Data: Often used by startups to compensate for the lack of real data, requiring strict data cleaning [10][20]. - Data Security Concerns: Increasing data security issues are highlighted, with incidents of unauthorized data transmission raising concerns about privacy and safety, especially as robots enter domestic service sectors [11][12]. - Benchmarking and Evaluation: The lack of a unified evaluation benchmark in the embodied intelligence sector is noted, with Stanford University introducing the Behavior 1K benchmark to assess embodied intelligence models, which could accelerate technological development [17]. Additional Important Content - Research and Development Efficiency: Companies are urged to optimize R&D processes and enhance cross-department collaboration to improve efficiency in response to industry demands [13]. - Physical AI's Role: Physical AI is recognized as crucial for simulation modeling, with applications in various industrial scenarios, showcasing its potential to enhance intelligent attributes [18][19]. - Software Ecosystem: The robotics software ecosystem comprises models, data analysis, simulation tools, and evaluation systems, attracting numerous tech companies to participate and create commercial opportunities [21]. - Future Trends: Over the next 3-5 years, the three algorithmic approaches are expected to coexist and evolve gradually, with hierarchical frameworks remaining relevant for industrial applications while VLA models gain traction in human-robot interaction [9]. This summary encapsulates the key points discussed in the conference call, providing insights into the current state and future directions of the embodied intelligence industry.