SIASUN(300024) -·2025-02-17 08:27Summary of Conference Call Notes Industry Overview - The discussion primarily revolves around the advancements in the fields of autonomous driving and humanoid robotics, focusing on the two main model types: Vision Language Model (VLM) and Vision Language Action Model (VAM) [2][3]. Key Points and Arguments 1. Model Types: - VTM (Vision Language Model) relies on large models for decision-making, converting commands into text instructions, and utilizing traditional motion control algorithms. It is noted for its real-time capabilities but suffers from understanding and cumulative errors [2][3]. - VRA (Vision Language Action Model) creates a closed-loop from visual and language inputs to action outputs, but is limited by computational power and currently only applicable in controlled environments [2][4]. 2. Performance Metrics: - VTM achieves an accuracy rate of over 80%, while VRA's accuracy is approximately 74%. The control frequency for VTM can reach 100 to 1,000 Hz, whereas VRA only achieves 3 to 10 Hz [5]. 3. Challenges in Transitioning Technologies: - Transitioning autonomous driving technology to humanoid robotics faces significant challenges due to differences in information transmission and the complexity of signal conversion in robotics [6][11]. - The input-output parameter dimensions in robotics are much higher, complicating the control mechanisms and requiring extensive data training [12]. 4. Data Requirements: - High-quality data is crucial for training models in humanoid robotics, with estimates suggesting that the data training volume needed is at least ten times that of autonomous driving [13]. - Current data collection methods include real data, which is costly and limited, and synthetic data, which is abundant but may not accurately reflect real-world complexities [15]. 5. Short-term and Long-term Strategies: - In the short term, VRM (a combination of rule-based and foundational control models) can accelerate the commercialization of humanoid robots despite some understanding errors [14]. - Long-term goals include achieving a fully integrated model (VRNI) that can operate across all scenarios, necessitating vast amounts of high-quality virtual and synthetic data [18]. Additional Important Insights - Current Key Players: - Notable companies in the humanoid robotics supply chain include Tesla and its related enterprises, as well as various domestic and international firms such as Zhongtian Technology and Ridi Intelligent Drive, expected to see significant changes by 2025 [20]. - Application Scenarios: - Humanoid robots are anticipated to find immediate applications in specific factory settings, such as assembly and welding, where high-quality data can be accumulated [17]. This summary encapsulates the critical insights from the conference call, highlighting the ongoing developments and challenges in the fields of autonomous driving and humanoid robotics.