According to Zhito Finance APP, Everbright released a Research Report stating that the balance and mobility of humanoid Robots have generally reached a relatively mature level. The core factor limiting the further expansion of the Industry is training, particularly the training required to achieve further capability generalization. Generalization training relies on a large amount of real-world data, which is costly to obtain directly through human demonstrations, and high-quality robot movement materials are relatively scarce. NVIDIA's (NVDA.US) Cosmos model can help developers generate exponentially scalable synthetic motion data, potentially ushering in the era of physical AI for Robots.

Event:

From January 7-10, 2025, CES 2025 will be held in Las Vegas, USA, where NVIDIA CEO Jensen Huang will deliver the opening keynote speech. The main points regarding Robots include:

1) The launch of a generative world foundation model development platform called Cosmos, introducing the 'Physical AI' concept that can generate high-quality synthetic data for training Robots and autonomous driving systems.

2) Leading the debut of 14 humanoid Robots, including Galaxy Universal G1, Star Motion Era Star1, Zhiyuan Robot Expedition A2, Fourier Universal Robot GR-2, and Xpeng's Iron Robot.

Comments:

The development cost of traditional physical AI models is very high, requiring a large amount of real-world data and testing. NVIDIA has launched the Cosmos model, which can generate physics-based videos based on inputs like text, images, and videos, as well as data from Drone sensors or motion. These models enable physics-based interactions, object persistence, and the generation of high-quality simulated industrial environments (such as warehouses or factories) and driving environments (including various road conditions), allowing Siasun Robot&Automation to more realistically simulate the motion and interaction of objects in simulated environments. The Cosmos model can collaborate with Omniverse, enabling developers to more easily produce large amounts of controllable and realistic synthetic data, creating a virtual world that adheres to physical laws, constructing Physical AI robotic systems that can train robots and autonomous driving, thus being more cost-effective than traditional data collection methods.

Cosmos has three versions: 1) Nano (approximately 15B, an ultra-low latency real-time model suitable for deployment on edge devices), 2) Super (34B, a high-performance baseline model supporting plug-and-play fine-tuning and deployment), 3) Ultra (approximately 70B, the highest in accuracy and quality, suitable for large-scale Datacenter scenarios). These models have been trained on 18,000 trillion tokens, including 20 million hours of real-world autonomous driving, Siasun Robot&Automation, Drone footage, and synthetic data, allowing AI to understand the physical world through training. Currently, companies like Galaxy General, 1X, Agility Robotics, FigureAI, Xiaopeng, and Uber have taken the lead in trialing these models.

Risk analysis: the risk of slow technological implementation, the risk of demand not meeting expectations, and the risk of the humanoid robot industry progressing slower than expected.