谷歌正在开发一部 “机器人宪法”，其目的是避免机器人更加安全，避免伤害人类。“机器人宪法” 实际上是谷歌的数据收集系统 AutoRT 中的安全护栏。AutoRT 利用了基础大模型的潜力，有助于开发能够理解人类实际目标的机器人。它通过收集数据帮助扩展机器人学习，从而更好地训练机器人适应现实世界。作为安全护栏的 “机器人宪法” 采用 “以安全为重点的提示”，指示大语言模型避免选择涉及人类、动物、尖锐物体甚至电器的任务。其灵感来自科幻作家艾萨克·阿西莫夫（Isaac Asimov）的 “机器人三定律”，即机器人不得伤害人，也不得见人受伤害而袖手旁观；机器人应服从人的一切命令，但不得违反第一定律；机器人应保护自身安全，但不得违反第一、第二定律。为了提高安全性，DeepMind 对机器人进行编程，如果机器人关节受力超过一定阈值就会自动停止，并安装了一个物理终止开关，让人类操控机器人停止工作。AutoRT 结合了视觉语言模型（VLM）、大语言模型（LLM）以及机器人控制模型（RT-1 或 RT-2），让机器人在新环境中收集训练数据。AutoRT 可一次性安全指挥 20 个机器人。谷歌部署了 52 个机器人，收集了包括 6650 个独特任务的 77000 次机器人试验的不同数据集。
Google is developing a “Robot Constitution” with the aim of making robots safer and preventing harm to humans. The “Robot Constitution” is essentially a safety barrier within Google’s data collection system called AutoRT. AutoRT leverages the potential of large-scale models to aid in the development of robots that can understand human objectives. By collecting data, it helps expand the robot’s learning capabilities, enabling better adaptation to the real world. The “Robot Constitution,” serving as a safety barrier, employs “safety-focused prompts” that instruct the large language model to avoid tasks involving humans, animals, sharp objects, or even electrical appliances. It draws inspiration from science fiction writer Isaac Asimov’s “Three Laws of Robotics,” which state that robots must not harm humans or allow them to come to harm, robots must obey human commands without violating the first law, and robots must protect their own safety without violating the first and second laws. To enhance safety, DeepMind has programmed the robots to automatically stop if joint forces exceed a certain threshold and has installed a physical stop switch for humans to control the robot’s actions. AutoRT combines visual language models (VLM), large language models (LLM), and robot control models (RT-1 or RT-2) to allow robots to collect training data in new environments. AutoRT can safely command up to 20 robots simultaneously. Google has deployed 52 robots and collected various datasets from 77,000 robot experiments involving 6,650 unique tasks.
微信扫一扫