Multimodal Embodied Attribute Learning by Robots for Object-Centric Action Policies
Autonomous Robots
Xiaohan Zhang, Saeid Amiri, Jivko Sinapov, Jesse Thomason, Peter Stone, Shiqi Zhang
Embedded Files
Intelligent robots are able to interact with objects through exploratory actions in real-world environments. For instance, a robot can use a look action to figure out if an object is RED using computer vision methods. However, vision is not sufficient to answer if an opaque bottle is FULL or not, and actions that support other sensory modalities, such as lift and shake, become necessary. Given the sensing capabilities of robots and the perceivable properties of objects, it is important to develop algorithms to enable robots to use multimodal exploratory actions to identify object properties, answering questions such as "Is this object RED and EMPTY? In this article, we use attribute to refer to a perceivable property of an object and use behavior to refer to an exploratory action that a robot can take to interact with the object.
Page updated
Google Sites
Report abuse