机器人大模型深度报告：我们距离真正的具身智能大模型还有多远？

Core Insights - The key to industrializing humanoid robots lies in overcoming the limitations of traditional industrial robots, which are based on deterministic control logic and lack perception, decision-making, and feedback capabilities [1] - The rise of multimodal large models provides humanoid robots with an "initial brain," enabling intelligent evolution and continuous improvement in model capabilities and product performance through a data flywheel [1] - Current intelligent models are still at the L2 initial stage, facing challenges in modeling methods, data scale, and training paradigms, with high-intelligence large models being a core variable in the path to general humanoid robots [1] Progress in Robot Large Models - The rapid evolution of robot large models is driven by breakthroughs in architecture and data [2] - Architecturally, models have progressed from early language planning models to end-to-end action output, integrating multimodal perception capabilities [2] - By 2024, the π0 model will introduce an action expert model with an output frequency of 50Hz, and by 2025, the Helix model will achieve a control frequency of 200Hz, enhancing operational fluidity and response speed [2] - The data structure now includes a collaborative system of internet, simulation, and real machine action data, with real machine data collection relying heavily on high-precision motion capture equipment [2] - The mainstream training paradigm is shifting from "low-quality pre-training + high-quality fine-tuning" to "data pile optimization," indicating a transition in model intelligence leaps [2] Future Development Directions of Large Models - Future embodied large models will evolve in three areas: modality expansion, reasoning mechanisms, and data composition [3] - The next phase is expected to introduce additional sensory channels such as touch and temperature, enhancing the robot's perception capabilities [3] - Architectures like Cosmos aim to provide robots with "imagination" through state prediction, creating a closed loop of perception, modeling, and decision-making [3] - The integration of simulation and real data for training is becoming the mainstream direction, with high-standard, scalable training environments being crucial for general robot training systems [3] Investment Recommendations - Companies to focus on in the model sector include Galaxy General, Star Motion Era, and Zhiyuan Robotics [4] - In the data collection field, attention should be given to Qingtong Vision, Lingyun Light, and Aobi Zhongguang [4] - For data training environments, Tianqi Co., Ltd. is recommended [4]