Google DeepMind Unveils AutoRT: Robot Constitution Ensures Safety and Prevents Harm
Google DeepMind, one of the largest tech companies, recently announced three impressive large language models (LLMs) aimed at enhancing the performance of robots. Among these models, AutoRT stands out as the most intriguing and innovative offering, as it introduces a Robot Constitution that enforces rules preventing machines from causing harm to humans. The other two models, namely SARA-RT and RT-Trajectory, are designed to facilitate robot training.
As we continue to integrate robots into our society, there is often a lingering fear of a future reminiscent of the Terminator movies, where killer machines dominate. To usher in a new technological age and ensure wider acceptance of robots, it is crucial to assure the public that these machines will prioritize human safety. In a positive development, Google DeepMind has taken a noteworthy step forward by introducing not just one, but three solutions to address this concern.
Let’s delve into the functionality of AutoRT, SARA-RT, and RT-Trajectory. AutoRT boasts safety guardrails in the form of a Robot Constitution, a set of prompts that guide robots in selecting tasks while prioritizing safety. Drawing inspiration from Isaac Asimov‘s Three Laws of Robotics, DeepMind has incorporated additional safety roles into AutoRT. For instance, the Constitution prevents robots from engaging in tasks involving humans, animals, electrical appliances, or sharp objects. However, Google emphasizes the importance of implementing additional precautions through programming to ensure optimal security. This might entail programming robots to halt automatically if their joints experience excessive force, thereby preventing accidents that could harm humans. Additionally, a human supervisor could maintain direct oversight of active robots and possess a deactivation switch as a precautionary measure in the event of catastrophic failures.
In addition to the Robot Constitution, Google has developed SARA-RT and RT-Trajectory. SARA-RT aims to enable robots to continue learning and maintain their performance as they acquire new tasks. This ensures that robots do not experience performance lag or crashes over time. On the other hand, RT-Trajectory allows robots to learn from human demonstrations. By observing human behavior, robots can learn various tasks, such as cleaning a room by watching someone dispose of trash in a bin. Furthermore, RT-Trajectory allows robots to learn from video uploads, further expanding their knowledge and capabilities.
While Google’s Robot Constitution focuses specifically on the integration of AI in robots, it is worth noting that artificial intelligence is utilized for a multitude of purposes. Consequently, effective training methods are crucial to ensure optimal performance from these AI systems. A recent innovative approach called Meta-learning for Compositionality has provided promising results. This method allows AI tools to apply different rules to newly learned words, offering feedback on adherence to these rules. Scientists have successfully demonstrated that neural networks can achieve systematic generalization, rivaling or even surpassing human performance in certain scenarios.
Google’s introduction of a large language model providing a Robot Constitution for robots worldwide is a significant milestone. Open-source in nature, this model has the potential to contribute to the safe use of robots by a broader audience. Additionally, the company’s development of SARA-RT and RT-Trajectory will undoubtedly simplify the training process for robot assistants, making it as simple as training a pet.
For further information on the open-source Robot Constitution, please visit DeepMind’s webpage. Stay up to date with the latest digital trends and tips on Inquirer Tech.
