LaST₀：让机器人拥有物理直觉，抛开语言拐杖像人一样思考动作

Core Insights - The article discusses the advancements in robotic intelligence through the LaST₀ framework, which allows robots to perform tasks more efficiently by utilizing a latent space for physical simulation rather than relying on explicit language-based instructions [5][8][28]. Group 1: LaST₀ Framework - LaST₀ enables robots to predict and encode future states directly in a compact latent space, focusing on visual dynamics, 3D geometric structures, and the robot's own bodily awareness [6][8]. - This framework enhances the efficiency and accuracy of robotic actions by eliminating the need for language translation, allowing for a more natural and fluid interaction with the physical environment [8][28]. Group 2: Dual Expert System - The LaST₀ framework incorporates a "dual expert" mixed Transformer architecture, consisting of a slow-thinking "reasoning expert" and a fast-reacting "action expert" [10][12]. - This design allows for real-time knowledge synchronization between the two experts, enabling robots to think and act simultaneously without delays [12][22]. Group 3: Performance Metrics - In rigorous testing across various simulated and real-world tasks, LaST₀ achieved an average success rate of 82% in the RLBench simulation, significantly outperforming previous models [14][16]. - The overall reasoning speed of LaST₀ reached 15.4 Hz, nearly 14 times faster than the language-based CoT method, which operated at 1.1 Hz [14][15]. Group 4: Generalization and Versatility - LaST₀ demonstrates remarkable generalization capabilities, allowing it to adapt to various robotic forms, from industrial arms to humanoid robots, by simply adjusting the end-effector's action dimensions [23][24]. - This adaptability indicates that the latent space physical reasoning ability is a universal skill, independent of specific robot configurations [24][28]. Group 5: Future Implications - The advancements presented by LaST₀ are expected to significantly impact industrial applications, enabling robots to handle tasks like assembly and sorting with greater fluidity and adaptability [28][30]. - In domestic and service sectors, robots will be able to perform complex household chores more safely and naturally, while in specialized fields like surgery and space exploration, they will operate independently in challenging environments [30][31].