Robot that seeks help from you when confused

2019-12-05

Artificial intelligence models always encounter ambiguous situations, making it difficult for them to react with just instructions. This poses a significant problem for autonomous agents, such as navigating through an apartment and facing the risk of being trapped by multiple paths, analogous to the typical "analysis paralysis" in humans. To address this issue, researchers from Amazon's Alexa AI department have developed a framework that empowers agent robots with the ability to request assistance in specific situations. The framework utilizes a model confusion-based approach, which asks questions based on the predefined confidence threshold to determine their level of confusion. The researchers claim that this threshold has increased the success rate of autonomous agents by at least 15%. In a preprint paper, the researchers wrote, "Imagine a scenario where you want a robotic assistant to place your wallet on the bed... There are two doors in the scene, but the preset instructions only tell it to pass through the doorway. In this situation, the robot would obviously struggle to determine which door to enter. After extensive exploration, we found a solution: if the robot can engage in discussions and seek help from the user, the ambiguity of the real-world situation can be resolved." The team's framework incorporates two agent models: Model Confusion and Action Space Augmentation. The former mimics human users' behavior in confusing situations, while the latter is a more complex algorithm that learns to ask necessary questions at the right time during the navigation process. The second model is further fine-tuned with human interaction data to familiarize the robot with the environment. In this study, when the robot gets lost while navigating a simulated house, it sends a signal saying, "I'm lost, please help me!" and seeks assistance. When the user provides answers to the robot's subsequent questions, the action space augmentation corrects the original erroneous trajectory and utilizes a feedback system to prevent the same mistakes in the future. The researchers compiled a dataset consisting of 21,567 navigation instructions, with an average of approximately 3,100 words per instruction. They then conducted experiments to evaluate the robot's success rate and number of steps, with the "success" criterion being a navigation error of less than 3 meters. If this newly proposed strategy proves successful, it could potentially inspire Amazon's long-rumored home robots and accelerate their development. Additionally, Amazon's strategy appears to be particularly useful in continuously learning scenarios, as household robots applying this strategy can continuously improve their performance in customers' homes.

RECENT AI NEWS

RECENT AI TOOLS