Amazon Wants A Home Robot That Asks Questions When It's Confused
Wed, April 14, 2021

Amazon Wants A Home Robot That Asks Questions When It's Confused

AI models get confused when presented with different paths / Photo Credit: Sarah Holmlund (via Shutterstock)

 

AI models struggle to respond when they encounter ambiguous situations, said Kyle Wiggers of VentureBeat, a technology news platform. That’s problematic when AI models are tasked with navigating an apartment because they can become stuck once they are presented with several paths. To address this, Ta-Chung Chi and colleagues at Amazon’s Alexa AI division created a framework that grants agents the ability to ask for assistance in certain situations. Using a so-called model-confusion-based method, the agents made inquiries based on their level of confusion. The levels were determined “by a predefined confidence threshold.” The researchers claimed that it can boost the agents’ success rate by at least 15%. 

The team wrote in a preprint paper, “Consider the situation in which you want a robot assistant to get your wallet on the bed … with two doors in the scene and an instruction that only tells it to walk through the doorway.” Given this situation, it is difficult for the robot to exactly know which door to go through. 

However, if the robot can discuss the situation with the operator, then the situational ambiguity can be resolved. The researchers’ framework employed Model Confusion, which imitates human behavior under confusion. They also used Action Space Augmentation, a more sophisticated algorithm that automatically learns which questions to ask at a certain time during navigation. Human interaction data is also utilized to fine-tune the second model to enable it to be familiar with the environment. 

When a robot navigating a home becomes lost during navigation, it says, “I am lost, please help me!” Then, the user provides the answer to the robot’s questions. The Action Space Augmentation corrects “originally wrong trajectories.” The feedback is used to prevent similar mistakes from occurring in the future. According to the researchers, the robot “adjusted dynamically to unclear and erroneous human responses.” 

The team stated that their proposed strategy was “substantially” more data-efficient than other “previously proposed pre-exploration techniques” that involve robots to explore a new environment by themselves. They concluded, “We are among the first to introduce human-agent interaction in the instruction-based navigation task.”