国金证券：具身智能迫近临界点 关注人形机器人四大潜在商业化场景

Core Insights - The report from Guojin Securities highlights the increasing integration of "Bots" in the CCTV Spring Festival Gala, indicating a growing trend in humanoid robots' commercial applications. The report identifies four potential commercial scenarios based on the differentiation between ToB/C customer segments and the generalization requirements of brain and cerebellum capabilities [1][2]. Group 1: Commercialization Scenarios - ToB + Brain Generalization: Scenarios such as guiding, shopping assistance, and patrols, exemplified by Xiaopeng's Iron robot in showroom tours and company front desk duties, require fluent communication and a friendly humanoid presence [1]. - ToB + Cerebellum Generalization: Specialized industry scenarios, as demonstrated by Yushutech's performance in the CCTV Spring Festival Gala, showcasing 16 robots performing complex movements with millisecond-level coordination, necessitating human-like motion control and swarm collaboration [2]. - ToC + Brain Generalization: Personal companionship scenarios, illustrated by the emotional engagement capabilities of robots in the "Grandma's Favorite" segment, focusing on language-based emotional value delivery and user data interaction [2]. - ToC + Cerebellum Generalization: Home care scenarios, as shown in the New Year microfilm, where robots perform tasks like walnut cracking and folding clothes, requiring fine motor skills and visual coordination [2]. Group 2: Advances in Embodied Intelligence - Breakthroughs in one-shot learning systems have been achieved, with Figure AI's Helix 02 enabling dexterous, long-term autonomous behavior across entire rooms, while Tesla integrates Optimus with a shared simulation model for advanced scheduling [3]. - Pre-training of brain-side models has surpassed the Scaling Law threshold, with Generalist AI's GEN-0 confirming that the activation parameter for embodied intelligence is around 7 billion, indicating that challenges are more data and engineering-related than scientific [3]. - The cerebellum side has seen significant advancements, with Skild AI's universal cerebellum demonstrating rapid adaptation capabilities across different robotic forms, achieving a 500-fold increase in training volume compared to traditional models [3]. Group 3: Data Integration and Training - True machine data is essential for embodied training, with UMI providing a low-cost, lightweight real data collection solution at approximately $400, consisting of 3D-printed components [6]. - Simulation synthetic data is being enhanced through platforms like NVIDIA's Isaac & Sim and domestic robotics training platforms, which have shown that diverse simulation data can help bridge the Sim2Real gap [6]. - Human video data is gaining traction, with advancements in Gemini 3.0/3.1 Pro improving screen understanding and abstract reasoning, which may accelerate robots' ability to comprehend the real world [6].