Quality-Aware Active Re-Observation for Constrained Eye-in-Hand Robotic Manipulators

Object visibility in a camera field of view does not guarantee that local depth is usable for robotic manipulation. This paper proposes a quality-aware active re-observation supervisory control framework for constrained eye-in-hand stereo manipulators. The framework elevates ROI-level stereo reliability from a post-processing variable to an admission condition in the observation decision loop. Instead of using a binary target-visible judgment, each observation is represented by evaluability, mean ROI risk, reliable ROI coverage, and an integrated quality score. The controller then accepts the current observation, triggers active re-observation, retains the initial result, or records a failure boundary. To avoid the cost of continuous viewpoint search and the impossibility of knowing post-execution quality in advance, an online proxy strategy based on Jproxy is introduced. Four candidate viewpoints are generated from the current observation and filtered by workspace, pose, joint-limit, and visibility constraints; the top-ranked viewpoint is then executed. Experiments are conducted on a four-degree-of-freedom eye-in-hand stereo manipulator platform. The results show that the method identifies observations in which the target is detected but depth is unreliable, and improves target-ROI observation quality and evaluability through a deployable online re-observation strategy. Specifically, compared with the initial observation, the online proxy increases the evaluable rate from 70.0% to 76.7% and the mean integrated quality score from 0.828 to 0.853; the independent manual offline reference reaches an evaluable rate of 93.3% and a mean integrated quality score of 0.916. Relative to the post-execution quality upper bound, the online proxy strategy achieves an effective quality-cost trade-off under limited pre-execution information. These results indicate that active vision for constrained manipulators should incorporate local depth reliability into online observation decisions rather than considering only whether the target is visible or centered. The proposed quality-gated active re-observation framework is applicable to eye-in-hand robotic inspection, pre-grasp verification, and local measurement, especially when the target remains visible but stereo depth within the task ROI is insufficiently reliable for downstream control.