3 6 Ke·2025-08-26 08:54Core Viewpoint - Tesla is shifting its training strategy for the Optimus humanoid robot to a pure vision-based approach, moving away from traditional methods like motion capture and remote operation, aligning with Elon Musk's belief that AI can learn complex tasks through camera input [2][5][8]. Group 1: Training Strategy Shift - Tesla informed employees in late June that it would focus on using video recordings of human tasks to train the Optimus robot, allowing it to learn actions such as picking up objects and folding T-shirts [2]. - The abandonment of motion capture suits and remote operation is expected to accelerate data collection efforts [2][6]. - The new training method involves using five internally developed cameras mounted on employees' helmets and backpacks to capture multi-angle video data, enhancing the precision of the AI model's environmental understanding [6]. Group 2: Comparison with Industry Standards - Traditional methods in the industry, such as those used by Boston Dynamics, involve remote operation and motion capture suits to train robots [3]. - Experts suggest that while video data can be useful, it lacks the direct interaction experience that remote operation provides, making it challenging for robots to translate video data into real-world actions [3][6]. Group 3: Challenges and Expectations - The training demands for Optimus are expected to be significantly higher than those for Tesla's autonomous driving systems, with Musk indicating that the requirements could be at least ten times greater [8]. - Experts emphasize the need for Tesla to develop a universal action library to avoid the inefficiency of training each action individually [7]. - The complexity of teaching the robot to understand and perform tasks based on video observation is highlighted as a significant challenge compared to single-task training in autonomous driving [8].