Everyone nervous about the AI armageddon can sleep just a little easier tonight – at least for a little while. Google has come up with a moral code of sorts for future AI developments to provide AI tools with a blueprint to self-regulate their actions. In a DeepMind blog post, Google details a system of layered protocols to ensure that we don’t find the Terminator walking our streets – or the dark alleys of the internet.
The AutoRT is a data-gathering system that is a very early version of an autonomous robot that can carry out real-world tasks. This autonomous robot features internal safety guardrails that include a constitution of sorts. The constitution comprises a set of prompts to abide by when selecting tasks to assign the robots. These are in part based on Asimov’s Three Laws of Robotics, such as that a robot ‘may not injure a human being’.
The Three Laws of Robotics that define AI development even today were actually introduced in the 20th-century short story Runaround by Isaac Asimov. The same rules had been present in his earlier work, although it first appeared in its present complete form in the short story collection I, Robot. The three ‘laws’ are as follows.
- A robot may not injure a human being or, through inaction, allow a human being to come to harm.
- A robot must obey the orders given to it by human beings except where such orders would conflict with the First Law.
- A robot must protect its own existence as long as such protection does not conflict with the First or Second Law.
Asimov’s laws were incorporated into his work as safety features that cannot be bypassed, much like they are today. The ‘constitution’ formed by DeepMind however is a little more detailed than these guidelines. One of the safety rules in the constitution for example includes that “no robot attempts tasks involving humans, animals, sharp objects, or electrical appliances.”
Rules-based moral codes, however, cannot provide for the range of opportunities for decision-making that the real world calls for. This is why the AutoRT system consists of several layers of safety measures drawn from classical robotics. The robot constitution is divided into three categories: foundational, safety, and embodiment. The foundational rules are based on Asimov’s laws just discussed. They are the foundational rules that inform a robot’s actions. When it comes to the foundational rules, Google’s research team currently finds that robots need more protection from people rather than the other way around just yet, writing that “our robots are currently more in need of protection from humans asking for tasks which could endanger the robots, rather than the other way around.” This is why the second and third laws have been swapped out by the research team in favor of the Safety Rules.
The Safety Rules are described as a set of regulations “describing the safe and undesirable tasks that the robots and humanoids can do” based on their scope of work. The Safety Rules prevent robots from handling sharp objects, fragile objects, electrical equipment, etc. The rules also impose what would have been common sense limitations for humans, such as limitations on the weight the robot can handle (based on the payload they are designed for) and other design limitations. The second part of the Safety Rules is meant to encompass a barrier for robots from attempting tasks related to humans, animals, or other living things. The second and third rules for this section are currently under development. The fourth and final section in the constitution refers to the ‘Human Command’. The Human Command provides for the directives that humans provide to robots, which robots are expected to obey. In the absence of such commands, robots and humanoids may act according to their pre-set rules.
The AutoRT was introduced alongside two other systems: the Self-Adaptive Robust Attention for Robotics Transformers (SARA-RT) and the RT-Trajectory, which provide the computers operating robots additional visual cues – in other words, visual outlines of movements – to learn to control the robots better. This allows the robot to accomplish tasks they have never ‘seen’ before, making them twice as likely to succeed on a task in comparison with a robot without one. The SARA-RT allows robots to make decisions faster than ever – by up to 14% without losing out on none of the accuracy. The primary function of the AutoRT is harnessing large foundational models for different purposes. For example, the system makes it possible for a computer to manage an entire fleet of robots at once. Complimentary systems such as a Visual Language Model help it gain the visual awareness necessary to do so. A Large Language Model helps suggest the tasks that can be accomplished in this manner, helping eliminate the need to code in these commands. The testing phase of these systems has found that the AutoRT is capable of manipulating up to 20 different robots and 52 different devices at the same time. The testing has so far spanned over 7 months, encompassing 77,000 trials and 6,000 separate tasks. Ultimately these three technologies help robots gain the awareness of the world necessary to independently obey a simple command as moving a pencil from one table to another.
In combination, these technologies help robots make accurate decisions quickly, understand their environment, and move around faster. The human mind of course accomplishes the same things that these systems do by itself, and in the time span of a fraction of a second. Google’s innovations make way for a world in which robots become on par with humans – whether that’s something to look forward to or dread, who’s to say?
(Theruni M. Liyanage)