马斯克新思路：Optimus或能靠看视频学会折衣服

Core Viewpoint - Tesla is shifting its approach to training its humanoid robot, Optimus, by focusing on a "purely visual" method instead of using motion capture suits and remote control, aiming to accelerate data collection and align with CEO Elon Musk's belief that AI can learn complex tasks through cameras [1][4]. Group 1: Training Methodology - The new training method involves recording videos of workers performing tasks, such as picking up objects and folding T-shirts, to teach the robot [1][4]. - This change follows the resignation of the project lead, Milan Kovac, with Ashok Elluswamy now overseeing the Optimus project [1]. - The previous reliance on motion capture and remote control limited data collection, as workers spent significant time managing the equipment [4]. Group 2: Technical Insights - The new approach utilizes a custom setup of five cameras worn by workers to capture their movements from multiple angles, providing detailed data for AI model training [4]. - Experts suggest that relying solely on video data may hinder the robot's ability to translate actions into real-world interactions, as physical engagement is crucial for learning [3][5]. - The strategy aligns with Tesla's existing methods for training its autonomous driving software, which also relies heavily on camera data rather than other sensor technologies [7]. Group 3: Challenges and Comparisons - Experts indicate that Tesla must develop a generalized set of actions for the robot to learn various tasks efficiently, as training on individual tasks could be time-consuming [5][6]. - The scale of data required for training Optimus is expected to be significantly larger than that for Tesla's autonomous driving systems, with Musk acknowledging that the training demands could be at least ten times greater [7]. - The complexity of training Optimus is viewed as more challenging than developing autonomous driving technology, as it requires both understanding video content and possessing the skills to perform tasks [7].