Newsletter
Google has been working on robotics for a long time, but a breakthrough might be on the way soon via the DeepMind team. That’s because the recent advancements in artificial intelligence paired with current and next-generation robotics could be a groundbreaking combination. To prepare for robots that might be able to act based on advice from AI large language models (LLMs), Google created a “Robot Constitution.” In essence, these base-level guidelines are intended to prevent robots from going awry and potentially harming humans.
The Robot Constitution is based on Isaac Asimov’s Three Laws of Robotics, which were introduced in 1942. These laws are that robots “may not injure a human being,” must obey orders from humans as long as they do not violate the first law, and must protect their existence as long as doing so would not violate the first law. Essentially, protecting human life is of the utmost importance. Everything else, from following orders to protecting itself, is secondary for robots.
Google’s robot guidelines add some practical elements, as well. For example, the LLM instructing these robots will not suggest tasks involving humans, animals, sharp objects, or electrical appliances. Even with these laws, Google understands that they are not absolute. That’s why it has introduced AutoRT, a system that uses AI for additional safety. It’s a three-pronged system that uses visual processing data in a Visual Language Model (VLM), an LLM, and a robot control model (either RT-1 or RT-2). All of this data and AI-assisted processing creates more safety measures, like one that will instruct a robot to stop if the force on its joints exceeds a certain threshold.
The state of Google robot construction with AI technologies
Google’s robots are becoming more advanced, but there’s still a lot of work to be done. As the Robot Constitution dictates, there’s a lot that robots can’t do, such as tasks involving humans or machinery. There is also still a lot of oversight required for robots to complete certain tasks. Plus, these robots are currently prioritizing function over form, and don’t have flashy designs.
Google rolled out 53 AutoRT robots to four office buildings as part of more than 77,000 tests and had various levels of control. Some were controlled by humans. Others followed a specific script, and some were using the RT-2 model. Human operators also had a physical kill switch that would force stop a robot from continuing, if the need arose.
AI will become a massive part of robotics. Currently, a robot can use a VLM to observe its surroundings. In addition to LLM to figure out what logical task to complete, and an RT model to execute the task. This all happens autonomously. Without the need for human input. It’s gearing up to be a breakthrough, but there’s still a long way to go before we see AI droids all around.