Answers are grounded in this paper and cite their sources.
In recent years, visuomotor policy learning has gained increased attention in the robotics domain. Ideally, such policies enable robots to generate actions directly from observations with high efficiency and accuracy. In reality, performance is significantly degraded when the robot encounters previously unseen environments (e.g., a novel position of an object to grasp, disturbances in the camera viewpoint). Several studies have attempted to deal with this problem by training the model on large-scale datasets, but constructing such datasets raises another challenge of data efficiency. To address these limitations, this work presents a data-efficient visuomotor policy framework via active vision. To be specific, the proposed framework employs an object detection algorithm and expresses the robot state and actions in terms of the camera coordinate frame. Provided that the camera viewpoint is properly adjusted by active vision, the visuomotor policy can manipulate objects at any position, even at positions not included in the training data. The experiments performed in both simulation and real-world environments with a humanoid robot and a dual manipulator system highlight that the proposed framework outperforms the state-of-the-art approaches significantly by enhancing the success rate for the pick-and-place task at unseen object positions from 45 to 91%.
No takes yet. Share an insight, caveat, or question.
Lee et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: