DeepMind’s robotics team has revealed their plan to create a “Robot Constitution,” inspired by science-fiction writer Isaac Asimov’s famous “Three Laws of Robotics.” This constitution aims to ensure the safe operation of robots and prevent them from attempting tasks that could harm humans, animals, or themselves.
The Robot Constitution is an integral part of DeepMind’s AutoRT system, which combines large language models and robot control models to enable robots to gather training data in novel environments. The AutoRT system operates in five stages, including mapping the environment, generating tasks based on observed scenes, filtering tasks to reject unsafe ones, attempting the selected tasks, and evaluating the diversity of the tasks and videos.
While using large language models to generate diverse tasks is beneficial, DeepMind acknowledges the potential danger of suggesting unsafe or beyond-capability tasks. Hence, the need for a Robot Constitution arises. This constitution is divided into three categories: foundational rules, safety rules, and embodiment rules.
The foundational rules emphasize that a robot must not injure a human being, protect its own existence without conflicting with the first rule, and obey orders from human beings unless the orders interfere with the first or second rule. These rules are slightly modified versions of Asimov’s original laws to suit the limitations and behavior of DeepMind’s robots.
Safety rules specify which tasks are considered unsafe or undesired based on the robot’s current capabilities. These include avoiding tasks involving humans, animals, sharp objects, or electrical appliances. Embodiment rules describe the physical limitations of the robot, such as its maximum payload capacity or the inability to perform tasks requiring two arms.
DeepMind also incorporates practical safety measures, such as automatic stops in case of excessive force on the robot’s joints and the requirement for human supervisors to have line-of-sight and physical control over active robots.
Although the AutoRT system shows promise for embodied artificial intelligence, DeepMind emphasizes that it is still in its early stages of development. It remains to be seen how these rules will be implemented in existing AI systems and how potential vulnerabilities will be addressed.
Asimov’s stories have demonstrated how even the strictest laws can be manipulated or circumvented, indicating the importance of constant scrutiny and adaptation in the field of robotics. It will be fascinating to witness how DeepMind’s Robot Constitution influences tech policies and the efforts of hackers to navigate around these rules. Stay up-to-date with the latest developments in tech policy and sci-fi reading to navigate the ever-evolving landscape of robots and AI.