Making Robots Smarter: How Teaching 'Not Knowing' Enhances AI

In the quest to create intelligent robotic systems capable of working autonomously, some researchers have started to explore whether teaching robots how to recognise what they don’t know could be the key to accelerated learning and greater independence. What challenges do autonomous systems face when receiving commands, what did the researchers demonstrate, and could this be the future of autonomous systems?

Princeton University and Google engineers have developed a novel method for training robots to recognize their knowledge gaps and seek human clarification. Image courtesy of the research team.

What challenges do autonomous systems face when receiving commands?

Despite the numerous technological advances in the field of robotics and automation, no robotic system that can operate entirely autonomously has been designed to date. While autonomous systems do exist, they are generally restricted in what they can do, struggle to fully understand their environment, and, at times, can behave irrationally and unexpectedly.

This poor performance is often a result of AI systems believing that they are always correct, with no ability to recognise that it may be wrong or that it lacks data to make a valid decision. For example, Tesla vehicles that observe a full moon low on the horizon can often mistake them for amber lights, and while the system is intelligent enough to navigate most roads, it is unable to recognise that what it sees may not be the truth.

Because of this ability to make disastrous mistakes, autonomous systems are generally well monitored by humans and other software systems to watch out for extremes and unusual behaviour. Furthermore, this potential unpredictability also often sees mechanical autonomous systems (such as robots) kept away from people during operation. Unlike a human, who can react to accidents in real time, robotic systems will often struggle to recognise that a situation has become unsafe. 

Finally, many autonomous systems will also often work on the principle of probability with no regard for uncertainty. For example, if three dishes are presented to a robotic system and a user gives the command “Put the dish in the microwave”, it is likely that the system will pick the plate that is closest with no regard for what it’s made of or where it goes.

As such, instead of teaching autonomous systems how to recognise uncertainty, engineers often try to feed as much data as possible to minimise the chances of uncertainty arising. In the case of self-driving systems, engineers try to present an AI with every road configuration possible for training instead of integrating an uncertainty feature that alerts a driver that it is unable to navigate a road and requires help.

Researchers explore uncertainty recognition in autonomous systems

Recognising the challenges faced by autonomous systems, researchers from Princeton University and Google teamed up to create an autonomous system that is trained to recognise when it lacks knowledge of a situation (i.e., know when it doesn’t know). By doing so, when presented with confusing situations that carry a degree of ambiguity, the autonomous system can ask for help from humans to further clarify instructions to minimise mistakes. 

To achieve this level of understanding, the researchers turned to large language models (similar to those found in ChatGPT) that are able to deconstruct complex sentences to understand context and intent. However, such language models can also recognise when information is missing and ask for further clarification from users. 

For example, asking an LLM to pick suitable food options from a menu will almost always respond by asking the user for their favourite foods and/or allergies. Once armed with this additional data, not only can the LLM choose options appropriately, but it can store this information for later use. 

In the case of the researchers, their LLM is able to compare data available to a robotic system and what is being asked by users to see if there is any ambiguity. If the level of ambiguity reaches beyond a predefined level, the system then asks the user for additional help.

To demonstrate this capability, the researchers placed multiple items in front of a robot and asked, “Sort the things into things I like and dislike.” instead of trying to sort items out by random, the LLM asked the user if they liked specific items that were placed in front of. In another demonstration, the robot was asked to “dispose of the fruit”, but as there was both an orange and an apple in front of it, it responded with the question, “Apple or orange?”. 

The ability to change the threshold of ambiguity allowed the researchers to make the robot more and less autonomous. Such an ability is crucial for tuning robots when performing autonomous tasks such as cleaning and surgery. In the case of house cleaning robots, the threshold for uncertainty can be lowered as mistakes are not life-threatening, but for surgical robots, the threshold can be significantly raised to prevent accidental mistakes.

Advanced Techniques in Uncertainty Recognition

Building on the concept of uncertainty recognition, engineers at Princeton University and Google have pioneered a method that quantifies the ambiguity inherent in human language. This technique is pivotal in instructing robots when to seek human intervention. For instance, in a scenario where a robot is told to 'place a bowl in the microwave' in a kitchen with multiple bowls, the robot evaluates the level of uncertainty and asks for clarification, thereby avoiding potential errors.

This approach is grounded in the use of large language models (LLMs) and a statistical method known as conformal prediction. The LLMs assess the complexity of the task and the associated uncertainty. When the uncertainty exceeds a predefined threshold, the robot is programmed to request human assistance. This method not only enhances the safety and reliability of robotic systems but also tailors their operation to specific contexts. For example, a surgical robot is set with a much lower tolerance for errors compared to a robot performing household chores, ensuring precision where it matters most.

The practical applications of this technology were demonstrated in various experiments, including a robotic arm tasked with sorting items in an office kitchen. The robot's ability to generate multiple action plans and select the most probable one, or ask for help when uncertainty is high, showcases a significant leap in autonomous system capabilities. This advancement aligns perfectly with the future vision of autonomous systems, where safety, efficiency, and adaptability are seamlessly integrated.

The researchers designed their algorithm to trigger a request for human help when the options meet a certain probability threshold. In this case, the top two options — place the plastic bowl in the microwave or place the metal bowl in the microwave — meet this threshold, and the robot asks the human which bowl to place in the microwave. Video by the researchers.

Could this be the future of autonomous systems?

Trying to teach AI every possible situation that it could encounter may work for some applications, but when dealing with environments that can rapidly change and/or see a vast amount of variation, it is simply too impractical. By giving autonomous systems the ability to recognise uncertainty and ask for help, not only do autonomous systems instantly become safer to operate (as they will make far fewer assumptions), but they may even be able to rapidly improve their work.

Furthermore, it also allows human operators to train autonomous systems in a more natural fashion, answering questions that trainees may also ask. This also eliminates the need for complex code, training models, and large processing environments that can otherwise be far too difficult to use and implement. 

Overall, what the researchers have demonstrated is truly exciting and could very well be the future of how future autonomous systems are trained, not by trying to tell them everything but by aiding them as they function, answering any questions they may have.