英伟达不能既当裁判员、又当运动员，Arm自然也一样。毕竟如果Arm亲自下场做芯片，高通、联发 科、小米、苹果等客户又该如何自处？技术授权商成为自己的竞争对手简直是"画美不看"，大家也都怕 Arm将新技术藏着掖着，因为一旦他们对最新的IP搞"捂盘惜售"，这些芯片厂商就得傻眼了。 Arm终于还是下决心亲自造芯了。日前Arm首席执行官雷内・哈斯在路透社的采访中表示，Arm公司已 完成的芯片名为Compute Sub Systems（CSS）的"物理载体"产品，正投资开发自有芯片，未来会打造从 Chiplet（芯粒）、实体芯片、主板到系统的 "全套产品"。 消息一出外界纷纷惊呼，"Arm莫不是疯了，居然要自己下场造芯片"。大家之所以会认为Arm想不开， 盖因Arm(Advanced RISC Machine)堪称是移动互联网时代最成功的处理器架构，作为低功耗、高性能处 理器的代表，Arm架构广泛应用于移动设备、嵌入式系统和物联网等领域的数十亿台设备中。 Arm的成功不仅仅是其卓越的能效比特质适合移动设备，更关键的原因是Arm公司采取了高度灵活的IP 授权模式。他们并不针对某个特定领域和场景设计及开发芯片，而是提出基础 ...

Summary of Conference Call on Robotics Industry Industry Overview - The humanoid robotics commercialization is still in its early stages, primarily applied in standardized processes within industrial settings, such as material handling in automotive manufacturing, but the actual usable scenarios are limited. Future applications are expected to emerge in standardized processes with high labor costs in hazardous environments [1][4] Key Points and Arguments - **Challenges in Commercialization**: Humanoid robotics face dual challenges in hardware and software. Hardware improvements are needed in actuator precision, sensor accuracy, power density, and battery life. Software improvements are required in human-machine interaction efficiency, multi-modal perception accuracy, visual image processing, and motion control stability [1][5] - **Data Collection Solutions**: To address the scarcity of training datasets, solutions include increasing real data collection (e.g., Zhiyuan's simulated living spaces) and employing physical simulation methods (e.g., NVIDIA's techniques) to enhance data quality and accelerate commercial application expansion [1][6][7] - **Training Data Efficiency**: By adjusting scene parameters or modifying scenarios, a small amount of real-world interaction data can generate hundreds to thousands of data points, significantly improving data acquisition efficiency and reducing costs. The future mainstream approach may combine real data collection with simulated data generation [1][8] - **Trends in Robotics Models**: The development of large models for robotics is trending towards multi-system architectures, such as NVIDIA's Grace Hopper. Future models need to address multi-modal and generalization capabilities, enabling robots to understand visual, linguistic, auditory, and tactile information [1][9] Additional Important Insights - **Technological Progress**: In the past two to three years, significant technological advancements have been observed in the humanoid robotics sector, with companies like UBTECH demonstrating impressive motion capabilities. However, humanoid robots still struggle with executing simple yet complex tasks, indicating that their intelligence level has not yet reached a fluid stage [2] - **Communication Protocols**: The EtherCAT protocol, with its distributed architecture, controls communication latency at the microsecond level, outperforming traditional CAN bus protocols and other real-time industrial Ethernet protocols, positioning it as a potential mainstream communication protocol for robotics [3][12] - **Market Developments**: DRECOM is set to release a new NPU and DMC stacked packaging product, suitable for running large models on the edge, expected to enter the market by 2025 or 2026. This indicates a growing focus on automation and data collection in investment trends [1][14] - **Sensor Technology**: The development direction for mechanical and tactile sensing is towards more precise perception and execution, enabling robots to understand real-world information accurately and perform fine operations [1][11] - **Chip Applications**: The current landscape for edge chips in robotics includes high-performance models from NVIDIA and Tesla for complex tasks, while domestic chips are being utilized for less demanding functions, indicating a growing opportunity for domestic chip performance enhancement [1][13]